The Unified Memory pool is what will continue to be the “game changer” in systems architecture, especially outside of data centers.
The reality is even cutting edge games and consumer workloads don’t actually take full use of the PCIe bandwidth of the GPU or the bandwidth of its GDDR memory. Even local AI use cases don’t substantially or meaningfully benefit from faster memory, at least to average consumers.
A unified memory pool does two things:
1) Lets systems optimize utilization based on need, rather than be confined to specific pools
2) Reduce overall memory cost, by letting system builders purchase a single type of memory in bulk instead of having to figure out GDDR vs DDR memory placement (important for SFF/portable machines)
So at a time when memory is expensive, unified pools make more sense. Even when memory becomes cheap and plentiful again, it’s just practical at this point to allocate a larger overall pool instead of managing discrete sets.
The one big drawback is security. A shared memory pool means side-channel attacks against memory from the GPU or CPU could potentially compromise the other as well, meaning memory-safe designs are going to be critical to security going forward (which is good for Rust adherents, I figure).
> The reality is even cutting edge games and consumer workloads don’t actually take full use of the PCIe bandwidth of the GPU or the bandwidth of its GDDR memory
Game dev here. For anyone reading this - it’s not because we’re lazy, it’s because _it’s really hard to do_.
One of the biggest differences between the current generation consoles and the current gen PCs is unified memory.
One related question that you need to follow that with is the associated costs of switching the whole studio to another engine that's technically better, or if proposing teach studio tailor-make their own engine the costs of that engineering, if presumably they have or learn the expertise to surpass whatever they're using currently.
I'm not a game developer, but it would also seem to be a link between resource usage by the engine, and whatever content the production side are making. For all the commentary about how brilliant the id software engines are, if you examine the levels you pass through they're also very efficient with what they demand out of the engine - it's like an orchestra playing well together, not one instrument that means you can do anything.
And conveniently, by making your machine non upgradeable, it allows the manufacturer to enforce market segmentation / charge a huge premium for small RAM upgrade (a la Apple)
LPCAMM2/SOCAMM2 exist, heck I think Framework is using LPCAMM2 in one of their new laptops.
Heck, I'm willing to bet that a lot of manufacturers would rather go that route than soldered in, if for no other reason than the relative cost of warranty work between the two.
However, people probably need to stop being obsessed with ultrathin laptops for that to happen.
I mean is it possible to make unified memory systems with good performance or is it not really feasible due to memory timing/trace length issues?
It’s possible if you’re willing to go with much slower RAM than GPUs like but CPUs often use. Thats what integrated graphics laptops have done for a long time right?
But can you get high end CPU and GPU performance with unified memory and maintain user upgradable memory in a reasonable way? Thats what I don’t know.
> I mean is it possible to make unified memory systems with good performance or is it not really feasible due to memory timing/trace length issues?
LPCAMM and similar solutions exist, but have never been demonstrated running at speeds that match what the leading soldered memory systems are using; there's always been some speed penalty. I'm not sure we've ever seen a system demonstrated using LPCAMM or similar for a 512-bit bus to match Apple's Max tier SoCs, so it's somewhat of an open question whether those solutions can offer upgradability at the high end of the market for unified memory systems.
Don't I/you wish. The mechanical junction adds no delay, only manufacturing expense, and the delay of purchasing new systems to keep up with OS bloat.
Actually the opposite is true. Socketed RAM can be made to overclock and adjust timings, while soldered ram, no. Two Lenovo's one soldered ( Carbon X1 ), one T590, one slot: Crucial 16GB, 260-pin SODIMM, DDR4 PC4-19200. Exact same processor, the X1 is DDR3 soldered on 532.0 MHz PC3-1066. The T590, has DDR4, PC4-19200, 1200Mhz.
Both have a Core i7 8665U... and the T590 is much faster, with socketed ram.
I think you'll find that in the current day, high speed LP(?)DDR5 requires a better signal path than what the SODIMM can provide. Which is why laptop makers initially moved to soldered RAM before moving to CAMM (probably only for the high end ones).
how about the LPCAMM route? Framework uses LPCAMM2 in 13 Pro laptop mainboards and claims that it satisfies the iGPU and NPU hardware without needing soldered RAM
AFAIK PCIe6 just started getting implemented in hardware last year... PCIe7 Spec was just released last year too...
PCIe6 is a much larger change than 'just bump up the transfer rate', the encoding changed too (on top of the new code length, it's no longer NRZ,) so everyone needed to design and validate both the new encoding block, negotiation, etc etc.
That said, I'm guessing PCIe7 will be a 'smoother' transition from PCIE6, i.e. we might see 7.0 products in 2027. That will theoretically get you ~240GB/sec, on an x16 link, or hypothetically a little less than the hypothetical max of a current Strix Halo. (I'm guessing however, that PCIe protocol overhead will make the difference larger.)
Don't really buy the economic argument. For 99% pf all workloads you need at least an order of magnitude more system memory than gpu memory.
Most systems barely need more gpu memory than what is required for video, browsing etc.
Just because we found a new usecase doesn't flip that on its head.
Besides, I want to keep doing what I'm doing today. So if I need 128GB today and my local AI needs 128 GB then I'd need 256 GB to keep doing the same work.
The argument rather seems to be that we shouldn't use such expensive memory on the GPU. Which might be true if you only want to do inference on it.
DRAM optimized for CPU usage looks very different from DRAM optimized for GPU usage. You are leaving a lot performance on the table when you have a unified memory architecture. It makes sense in some situations, but it is not a silver bullet.
The "one big drawback" is the lack of consumer upgrades, and the seemingly arbitrary prices charged by vendors for memory upgrades at time of system purchase. I'm not saying it has to be that way, but seems like it has been so far :-(
That was the main reason for the big hype around Memristors 15 years ago. High density, high speed persistent memory to completely remove the need for hdd/ssds, potentially even removing the need for external memory altogether. So frustrating that it still seems like we're a long ways from that becoming reality. There's some renewed interest in Memristors as they can simulate neural network connections in models, so maybe the funding will return for it.
The one example of persistent memory that managed to reach the mass market was Intel Optane/3dXPoint (still popular today among people looking to save on RAM costs) and that used a kind of phase-change memory, which is but tangentially related to memristors. ReRAM is somewhat closer, but it's also been less successful so far.
Well, back in the day... The MacIIfx had video memory, ( dual ported ram ) that could be read and written to out of different ports. Wicked fast. It 486DX2s more than a year to catch up.
Memory safety is orthogonal to side-channels, and hardware-enforced isolation (e.g. IOMMU) is more powerful than compiler-enforced isolation (but both are good!)
I'm not sure everyone uses the terms consistently, but the difference is that the old "shared" memory was reserving a section to act as VRAM under the control of the GPU, ignored by the OS. The CPU ran the same kind of code pretending there is a "bus transfer" between host memory and graphics memory.
In unified memory, all the memory is host memory and data can go from program to GPU with zero copy movements. The addresses of buffers can be shared via appropriate MMU translation support, so that the application and graphics subsystem are communicating effectively through the basic RAM cache coherency protocols over the same buffers.
Edit to add: Aside from the zero copy transfer potential, it also means dynamic allocation strategies can shift the balance between host and graphics allocations on the fly. Individual image and message buffers can be allocated on the fly instead of setting a static split between the two worlds.
For these in specific, they appear basically transparently to the GPU. There's a lot of software/firmware stuff for this, but also a different hardware architecture - while the RAM is on the CPU die, the nvlink-c2c gives it extremely low latency and 600GB/s bandwidth between the GPU and CPU.
Shared memory of the past meant reserving a part of the memory for the GPU, which could then not be used or accessed by the CPU. If the CPU wanted to access something, it had to copy it from the GPU's section of the memory to its own. Unified memory means both just fully share the same memory.
No. Let’s define terms, as others have pointed out they’re not perfect.
Unified memory is what Apple is doing, other phones do, and many low end built in GPUs have done in PCs for ages. There is only one physical memory pool. Both the CPU and GPU can access it at full speed.
This means no copying between pools of memory. No speed penalty accessing the CPU memory from GPU or vice versa. If the GPU only needs 2 GB to draw the desktop it only uses 2 GB of the pool. Or it can use 45 GB if it needs it and the CPU doesn’t. But all memory has to be the same speed, and that ain’t cheap given how fast GPUs like things. I don’t know if expandable memory is possible, and they use the same bus do they compete for bandwidth. Seems theoretically easier to program for to me.
The opposite is what’s been common in graphics cards since the 2D era. CPU and GPU have their own memory and can talk over PCI/AGP/PCI-E. This is what I think they mean by shared memory, if it’s not what’s the point in touting unified?
In this model if the GPU uses 2 GB of its 12 GB total, the other 10 isn’t available to the OS at full speed and I’m not aware of any operating systems that would use it for programs/cache by default. If the GPU needs 45 GB… too bad. You have to page things in and out of GPU memory over the much slower system bus. Starting a game means loading assets into main memory then transferring them to the GPU (newer tech can accelerate this). But the CPU can have slower memory than the GPU saving money. Memory expansion on the CPU side easy. And the CPU saturating its memory bus has no effect on the speed of the GPU memory bus because it’s physically separate. More complicated memory model but it’s the one everyone uses used to.
Which is better is a matter of opinion and workload needs.
Yes, I know there is an actual difference vs. dedicated GPUs with their own VRAM. I say it's marketing because Apple popularized the unified memory term even though, as you said, it existed in iGPUs long before Apple Silicon and was called shared GPU memory.
> I don’t know if expandable memory is possible
It technically is. These new systems (mostly) get their high bandwidth by using more channels (wider bus) of normal RAM modules. A system that has LPCAMM2 sockets should allow using the same LPDDR5X memory but you'd need a socket per two channels. A typical PC only supports two channels so having four (two sockets) would double the bandwidth.
System RAM has much lower bandwidth and less predictable access. Notably, the transfer from system to GPU is very slow. About 30x slower. LLMs aren’t designed to queue or parallelise operations to account for this. They just become much slower.
Unified memory is only a feature because NVidia so aggressively uses VRAM for market segmentation.
The 5090 ($2k MSRP but realistically $3-3.5k) is almost the same as the RTX 6000 Pro (~$10k). Same memory bandwidth (1800GB/s). Slightly different CUDA cores (21k vs 24k). Big difference? VRAM (32GB vs 96GB).
NVidia ultimately doesn't want to upset this segmentation so the RTX Spark will never undermine their other offerings. This is why I think Apple has a real market opportunity if they choose to embrace it.
To this day I do not get why Intel doesn't just offer massive memory options for their cards. Just charge what it costs to add the extra memory, no upcharge, and they will never be able to keep up with demand. Cheap VRAM is enough to justify a lot of open source investment into challenging CUDA.
They took longer than everyone expected and then shortly after release they made announcements that made people worry that Intel might kill the project the way they tend to kill GPU projects.
That’s not massive, though. Make it 96GB at $2,000 (ok, probably impossible right now, but they could have before the surge in prices) and you’ll see developers work really hard to make AI tooling work for their cards, CUDA be damned. The same goes for AMD.
It’s like they both want to rely on market segmentation for VRAM too but fail to realize that it’s their only potential inroad right now.
If you buy three 32GB GPUs, that's 96GB total at a very reasonable price. An AI model splits easily by layers, so running on multiple GPUs is quite feasible.
Even low-VRAM cards are actually very useful for running the comparatively smaller dense layers in large local MoE models. This only requires transfering very small amounts of data across the PCIe bus (similar to pipeline parallelism) so it fits nicely around the existing bottlenecks on that hardware.
I have so many questions… Since Apple already sells unified memory systems, what is the market opportunity you envision? Do you see Nvidia and Apple as competitors, and how? (And I’m not suggesting they’re not, necessarily, but I want to hear where you’re coming from, and they do have very different markets.) Hasn’t Apple used storage size (RAM & disk) for market segmentation for decades? And how does a machine with 128GB unified mem not potentially cut into some people’s reasons for wanting a 96GB GPU?
I'm not the person you're replying to, but I wholeheartedly agree with them...
Quick background: doing AI inference requires three things. Lots of memory, lots of memory bandwidth, and of course plenty of compute that has access to that memory.
Quick reference: nVidia 5090 has 1,792 GB/sec bandwidth. 3090 gets about 1000 GB/sec. DGX Spark and AMD 395 whatever get about 275 GB/sec.
Apple M1 Max gets 400GB/sec, M5 Max gets 614GB/sec. Ultra variants get 2x that bandwidth, base variants get 1/2 that bandwidth. However... their compute is rather weak.
Right now, Apple's offerings are juuuuuust fast enough to run dense 27B models at usable speeds at like, 10% of the performance/watt of nVidia. They're world-leading general purpose CPUs but not killer GPUs.
By all accounts, these Windows PCs nVidia is touting seem to have DGX Spark like performance, which is less than impressive. Same with the upcoming AMD AI-oriented consumer stuff.
The other context here is that running your own AI at home is just starting to become feasible in terms of open model availability and the ability to run it at usable speeds. Many are interested in it for reasons of privacy, security, and cost certainty vs. buying tokens.
Since Apple already sells unified memory systems, what
is the market opportunity you envision?
nVidia and AMD can't make their consumer offerings too good at AI, because that risks interfering with their higher-margin data center sales.
(And, let's face it. Even if nVidia did release a 6090 with 64-128GB of memory for an affordable price, consumers wouldn't get their hands on them anyway because people would just start filling data centers with them)
So.
Now you see Apple's opportunity, right? No data center sales to interfere with. No relationship with nVidia or AMD to worry about.
They could choose to make an absolute beast of a home AI machine. The M5 Ultra, if announced, might be that. It's admittedly a niche market, but people are already buying 64GB+ Macs faster than Apple can make them and they're fetching high prices on the used market as well.
The only real questions are if this market is even something Apple would find time to care about, and if they could secure enough DRAM to make a go at it. They are enormous obviously but they're feeling the RAM pinch just like everybody.
They use different technology for their VRAM though. Apple, AMD Strix and NVidia DGX/RTX Spark use LPDDR, whereas discrete cards will be either GDDR or HBM. That directly impacts the memory bandwidth figures. As for compute available, Apple and AMD still have very good figures there for what's essentially a general-purpose iGPU that ships as part of the stock system, rather than a special-purpose piece of dedicated hardware.
Even if a Mac isn’t the fastest in raw numbers it may be faster if it can load the whole model in its ram (went up to 512 GB before shortages) than a couple 32 GB cards could with the data having to be constantly loaded over PCI-E. Because unified memory means the Apple GPUs can access all 512 GB at full speed.
My understanding is this is the advantage that’s pushing huge Mac Studio demand. Because it was the only way to give GPUs so much memory at price points anywhere near.
Yeah you can do way better once you’re in the 5 digits. But below that Apple had a specific advantage for some.
Apple offers relatively affordable options for a high-memory workstation that uses unified memory. They previously offered 256/512GB Mac Studios (both discontinued). Because of this they can keep larger models in memory.
BUT you just can't compete with NVidia performance for LLM workloads (mostly inference) for two reasons:
1. The memory bandwidth just can't compete with a 5090 (1800GB/s). The best current Mac is ~900GB/s. That directly caps tokens/sec and might be manageable but there's another problem; and
2. The raw FLOPS just can't compete with even a 5090. It probably needs to natively support FP4/FP8 to at least maintain a number format parity with NVidia. But beside that, NVidia just has more raw FLOPS.
According to Google, an M5 Max does ~70 FP16 TFLOPS while a 5090 does 380. If Apple can close that gap to at least be competitive and also hold larger models in shared VRAM, that would be a competitive advantage and it would directly attack NVidia's market segmentation.
The Mac Studio last came out March last year. So we may get an update in Q3. Many are pinning their hopes on this. But it might not happen until next year. When it was released the M4 was the state of the art and it came with either the M4 Max or M3 Ultra (which, as I understand it, is basically 2 M3s stuck together, kind of). What people are hoping for is an M5 Ultra with >1000GB/s of memory bandwidth, ideally 200+ FP16 TFLOPS and hopefully FP4/FP4 support.
You can chain Mac Studios together into a cluster with TB5 too.
But it's reasonably likely that the next Mac Studio will be only incrementally better than the last generation.
Not true. This is aimed squarely at the Strix Halo and Mac markets. It's basically just strictly better than the Strix, and it's not clear cut vs that Macs in any sort of blanket statement.
My M5 Max 128gb MBP decodes faster than one of my Sparks, but the Spark's prefill is so much faster it can often answer the same query before the mac's prefill is finished. If you have large prompts, low cacheability, etc., a spark might be a very good options.
Not to mention you get can get two sparks and the MBP will be 85%+ of the cost at half the RAM.
I'm kind of tempted to pick one up. Leave running big models to my dual dgx setup, and all the misc. random stuff on an rtx.
Prefill will be a huge deal if batched unattended inference of SOTA models (on consumer platforms) becomes viable, because at that point it's the main remaining bottleneck. If running 30 inferences together boosts your decode throughput to 3x (that's consistent with some very rough experiments, though these haven't even looked at trying to mask SSD offload latency just yet), that's a 10x in total decode time but a 30x in total prefill time, because prefill workloads are fully compute bound already on consumer platforms and don't benefit from batching much at all.
yeah, you only see double digits in performance degradation from going from pcie 5 to 3 with a 5090 (at x16 speed), with everything else its like in the single digits area.
And the thing we gamers forget is that we’re the outlier. We’re the edge case.
Most consumers will never really care about, let alone see, the difference in PCIe or memory bandwidth impacts from such a shift to unified memory pools. We might (being, at least in my case, a huge nerd), but I’m increasingly of the opinion that if modern blockbuster games are built for upscaling/reconstruction anyhow, then suddenly such sacrifices to performance seem acceptable relative to the gains in efficiency.
Well I mean, the idea with games is it all fits in vram. You really don't want to be thrashing. It's that things are still so slow that they must be avoided entirely, no?
No copy unified memory will help with that but you do pay the read speed costs.
It’s also the reason, why you will never be able to repair or upgrade your computer in the future. From technological point of view these are indeed big advancements.
However, I couldn’t care less about faster CPU when:
Intel was doing UMA with their i740 graphics in the late 90s. Codename TIMNA was cancelled, but they pioneered it and used it on their you/cpu chips as well as their breakthrough 810 chipset that dominated graphics market for a decade. It was despised because it wa ubiquitous and a low performing graphics engine but games had to accommodate it.
FWIW, the O2's UMA let it handle far more textures than almost any other contemporary system with reasonable performance.
Most other SGIs had single or low double-digit megabytes of texture memory, whereas the O2 could host one gigabyte of unified memory and use a huge chunk of that for textures.
I want unified but not uniform - everything can address anything, but you can add slower RAM to the system without requiring an entirely new chip. NUMA is cool.
Everything that doesn't have a discrete GPU has unified memory these days. If you're asking for something closer to the RTX Spark or Apple Silicon then look at AMD's Strix Halo systems.
As a Rust adherent, please do not put words in our mouths or set up unrealistic expectations for other people by linking together concepts at a very shallow level.
Language level memory safety has no answer for hardware security flaws which is what side channel attacks are. No programming language can provide memory privacy if another chip in your machine can read your memory. Just like no programming language can protect your application from a kernel vulnerability of the kernel it’s running on.
Damn. That wasn’t my intention at all, I was just pointing out that Rust has another reason to see wider adoption vis a vis the usual Valley advertising bullshit of deliberately conflating hardware security with software security. I personally give no fucks what something is written in, only that it’s written well enough that I don’t have to twist arms or babysit yet another sloppy piece of code in my enterprise.
"I am not sure how many people will run AI models locally. It still seems like a niche application to me. However, it will make decent machines to play video games."
I don't know who will be the winner but with some of the recent releases from gemma it seems more probable that you may run some models locally if only from a cost perspective, not even considering business security. Not sure how this type of architecture would make for good gaming though, puts into question the whole statement.
"Ranked in the top 2% of scientists globally (Stanford/Elsevier 2025) and among GitHub's top 1000 developers" - side note but this guy puts this everywhere, gives me probably the inverse of what he is marketing for.
"I am not sure how many people will run AI models locally. It still seems like a niche application to me. However, it will make decent machines to play video games..."
This is the 2026 edition of Ken Olsen:
"There is no reason anyone would want a computer in their home"
> This is the 2026 edition of Ken Olsen: "There is no reason anyone would want a computer in their home"
Digging into this:
> In conclusion, there is evidence that Ken Olsen did doubt the need for computers in the home, but the evidence is based primarily on the testimony of David Ahl who was perturbed when the personal computer project he championed at DEC was not supported by Olsen in 1974.
> Olsen’s resistance may have been similar to that expressed by another DEC executive, Gordon Bell. In 1980 Bell thought home terminals would act as gateways to remote computers which would provide appropriate services.
It was supposedly said in 1977: most computers at that time were not small, and so it would not be surprising that people would not expect the general public to desire a large, power-hungry, noise-y apparatus in their house.
This is why I'm bearish on Anthropic, OpenAI, and friends. I am not confident that we will continue to see the same pace of improvement in frontier model capabilities as we have seen over the past year or two - not using similar mathematics at least. But I think that getting results that are close enough to the same standard to be a realistic substitute but in a model small enough to run locally may well happen quite quickly. And if it does - where is the moat to defend these AI organisations with their astronomical budgets when they're already starting to price more realistically and that's already killing a lot of the hype they've enjoyed until very recently? They have an accidental moat because they bought up the global supply chain for storage but that surely isn't going to last once the data centres to hold that storage are becoming liabilities.
If model performance asymptotes and CPU/GPU and RAM keep growing, even slowly, then eventually we will have frontier models on desktop that are totally competitive with hosted. It’s only a matter of time.
You already can if you’re willing to spend many thousands of dollars on a beast of a machine. I’m talking about middle tier desktops and laptops here. Maybe eventually even phones.
The only way hosted stays strongly competitive in that world is if they can keep pushing the frontier or by playing the classic social media and SaaS games of network effect building and integrations.
Many people might still use hosted, of course, but what I really mean is that their multiples won’t be justified and they will have little to no moat. AI will become commoditized, like a sophisticated next generation form of an encyclopedia with search.
We kinda ended up with terminals connected to mainframes anyway. The terminal being the web browser, and the mainframe being SaS. So it wasn't that far off.
People take these quotes out of context all the time. Said in a business context, there was no need, at that time, for someone to have a personal computer.
There's no business justification in 1977 for a personal computer department at a business. It's similar to the gates quote about RAM (I think it was 64KB?).
These statements aren't meant to be forever quotes. Their business plan quotes.
That exact quote? No, never.
He said something like: current computers at the time had 64kb of RAM, so the OS was designed with a limit of 640kb, and he believed this would give them 10 years of future proofing. As it happened, that limit was reached much faster, in about 6 years.
He had a long career and presumably many successes, and is fallible like the rest of us. But a half-remembered zinger with no context makes for zippier posts I guess.
The early popularity of Minitel, the continued popularity of ssh/tmux, and the web browser itself indicates that bespoke client applications are not the only way. He wasn’t directionally wrong.
I will not be spending thousands in hardware to run the worlds most mediocre llms at meh speeds. Sorry. I know for llm bros they think every output made by an LLM is magic, like every NFT guy thought every NFT collection was game changing, but there's nothing useful you can do with llms and 128gb of RAM (and there never will be) unless you have llm psychosis. Who cares.
Nobody ever said that, at least not as an assertion or prediction. The actual instances of similar language are from multiple people describing their earlier thoughts before they learned it wasn’t true.
It’s better, it’s useful even for those who don’t have a deep knowledge of computers. I’d expect more AI users than programmers, than ms-word users, than excel users.
Local models aren’t deterministically equivalent in capabilities to foundation models. Home computers are turing complete; just like a mainframe. They are just slower. Often not slower enough to matter.
Most people are ok with slower. An AI that lets you edit a family picture, in say 30 seconds, locally is preferable to one that is instantaneous but requires you to submit that picture to examination/storage/training/sale in someone else's AI ecosystem. If i want to crop my ex out of family photos, i should not have to first give that photo to Microsoft. If want an LLM to write a book report for me, i dont want it also alerting my school. And if i write a memo for a client, and i want an LLM to check the spelling, i dont want that memo leaked either.
It’s completely technically possible to have cloud services where customer data is opaque to the provider. Some of Apple’s services are like this already, for example.
I think there’s a sweet spot currently with munging your data blindly on the server so that your client device battery still lasts all day.
Meanwhile Apple and others push on with making client side models more efficient so that eventually the server costs and complexities go away.
I'd like to think so but the existence of Google and Apple and Microsoft's cloud based photo tools with phone integration suggests that's false.
You could run a pretty good home server on $50 of gear and yet we never saw any real adoption of OwnCloud/NextCloud style products as an alternative to Google Drive/Photos or Apple Cloud.
Why should LLM/Transformers be any different? Especially when you need a proper expensive GPU to run them instead of a Raspberry Pi?
> Most people are ok with slower. An AI that lets you edit a family picture, in say 30 seconds, locally is preferable to one that is instantaneous but requires you to submit that picture to examination/storage/training/sale in someone else's AI ecosystem.
Maybe if you ask them that question, but if you show them two products, they'll definitely prefer the faster one. 30 seconds is a long time to watch a progress bar.
Fast and public, or slow and private. Not everyone wants, or is allowed to, share their data with the AI world. And do not doubt that every bit shared with an AI service will be used for training.
The question here is about markets though. Not everyone wants x but if the vast majority of people want y, x is going to be niche and expensive.
You don't think the commercials of Google's AI photo features aren't going to have an impact on Apple users of their phones can do a worse version of that feature and it takes longer?
Plus there's the other question. If this thing is slower ... what's the price? The desktop/mini-pc version of this is $3000, after all. At this performance level what is an acceptable price for the laptops?
People definitely aren't going to accept more expensive + slower ...
He’s just a braggart. When you see something like this in somebody’s personal bio on social media, it’s basically a banner that means “take everything I say in the context of me promoting myself.”
Qwen 3.6 is far ahead of Gemma for most (but not all) things. I've deployed it out across a number of M5 MacBooks and it's genuinely useful for many tasks. It won't replace an Opus or current gen Sonnet sized model but it's still amazingly good for its size and probably as good as or just a bit before Sonnet 4 era. Far more reliable for tool calling, coding, agentic tasks and faster than the Gemma models especially with MTP.
Qwen 3.6 is a toy compared to DeepSeek V4 Flash or Pro. These models can now run on Apple Silicon hardware with as little as 32GB RAM for the Flash (with 2-bit quant, which is still quite capable) using SSD offloading, with just-about-reasonable performance for interactive use, and far better performance on longer contexts than Qwen (due to the more efficient KV cache/attention mechanisms in DeepSeek).
Very significant improvements may be viable for unattended inference via large-scale batches, which can reuse sparse experts and thereby mask some of the latency involved - this is quite unique to DeepSeek, again due to its efficient KV cache.
1. Deepseek V4 is still in preview (training is not finished)
2. Qwen is much more demanding and borderline unusable on consumer hardware because it's a dense model. The 27B parameters are active all time for each token. It's not a MoE architecture where a router activates only some of them.
Deepseek V4 Flash still has 13B active params though? That is about half as many as Qwen3.6-27B (and much more than Qwen3.6-35B-A3B). Given that RAM (even on a base M4 or 'regular' Intel/AMD system) is like an order of magnitude faster than an SSD, even Qwen 27B running from RAM will be much faster than any Deepseek V4 model with SSD offloading. And the MoE will be much faster still.
Qwen 27B is also small enough to completely fit in a high-end consumer or mid-end pro GPU, like an RTX 5090 or Radeon PRO R9700. I found results claiming 30 tokens per second generation for 27B(-Q4_K_XL) on an R9700. I doubt you get more than 5 tokens per second doing SSD MoE streaming.
Even for relatively short contexts, I honestly already find the ~30B class MoE models to be only borderline acceptable in terms of speed on my laptop (Ryzen 7 7840U, 64 GB LPDDR5-6400), though I use Gemma 4 26B-A4B more than Qwen3.6 35B-A3B.
> even Qwen 27B running from RAM will be much faster than any Deepseek V4 model with SSD offloading.
If you have reasonable amounts of RAM to cache the most likely experts, that's not true at all. Qwen 27B is marginally faster on a nearly empty context, then falls behind as context length increases due to the different attention mechanisms. Prefill for Qwen is much faster, but you're still comparing vastly different model sizes and capabilities. DeepSeek Flash is the best deal overall.
> completely fit in a high-end consumer or mid-end pro GPU
Or you could fit the dense portion of a much more capable model and still take advantage of that hardware.
Is that how MoEs work? I though that an important constraint for MoEs is that experts need to be uniformly used to make sure they can be used effectively. If there is a 'common subset' that, if anything, sounds like a symptom of undertraining (i.e. the same trick will not work as well for Deepseek V4.1).
Also, even if your MoE hitrate is 90%, you still spend half your time waiting for the SSD, giving similar total speed to a 27B model!
Finally, it looks like Deepseek V4 is pretty much only runnable with antirez's ds4, and SSD streaming only works with Metal; but I would like to try what you say with llama.cpp which uses mmap to also potentially do SSD streaming. (I can maybe try the large Qwen3.5 MoEs?)
> as context length increases
What kind of context length do you consider reasonable, though? From what I know, all models (even frontier ones) start degrading once you pass a few hundred thousand tokens. So realistically, limiting context size might even improve quality, especially if you use token-efficient harnesses.
> Or you could fit the dense portion of a much more capable model and still take advantage of that hardware.
Your point about consumer hardware was that it would be "borderline unusable" when running Qwen 3.6 27B. However, you need much less hardware to run a 27B than DSv4 Flash. In addition, you can do the same 'trick' with low-end GPUs and small MoEs: my desktop with 32 GB DDR4-3200 and an RTX 2070 8GB can run the ~30B class MoEs at 20-30 tokens per second and similar speeds to my laptop.
For any given workload/session? Empirically, yes, that's what has been found across different models. There's quite a bit of predictability that makes caching helpful.
> Also, even if your MoE hitrate is 90%, you still spend half your time waiting for the SSD, giving similar total speed to a 27B model!
There are ways of masking some of that latency, though it requires some architecture-specific cleverness which is less directly applicable to a generic engine like llama.cpp.
> Finally, it looks like Deepseek V4 is pretty much only runnable with antirez's ds4, and SSD streaming only works with Metal
The llama.cpp folks are working on adding support, and the ds4 project is working on CUDA support for streaming inference, targeting the DGX Spark.
> From what I know, all models (even frontier ones) start degrading once you pass a few hundred thousand tokens.
DeepSeek V4 seems to do quite well on recall tasks even with large context. That's one plausible benefit of its compressed attention mechanism, compared to earlier models. Some degradation will likely still be there, but it's not necessarily obvious.
As for why people are calling Qwen 27B "borderline unusable" that may have to do with it being a dense model which makes for an increased compute intensity and pushes users towards discrete GPU platforms, since those tend to have the most compute overall as far as consumer hardware is concerned. I might agree that Qwen 27B is quite ideally tailored towards these platforms, but that does come with some limitations.
I have to disagree with most claims. I run Qwen3.6-27b at 260k context and 40-60 tok/sec. It handles most coding problems as well as Sonnet 4.6 under OpenCode on our production tasks. (As an experiment, I run the same prompts for the same issues in parallel for Qwen 3.6 and Sonnet 4.6 and usually see little difference in performance). I see zero degradation from quantization in practice.
Last time I tried running large MoEs on this PC, they had inferior quality at 2-3 bits compared to much smaller dense models at 5-6 bits, and were slower anyway.
A 260k context (close to the stock maximum for Qwen, though it's possible to extend it) will take ~16GB RAM for storing the KV cache, barring quantization tricks which severely degrade quality. That's a whole lot more than what DeepSeek requires for a similar context length, and makes it infeasible to batch multiple inferences together. This used to be the status quo for consumer inference, in fact it still is for models like Kimi and GLM (which can sometimes be smarter than even DeepSeek V4 Pro!) but we can also do better nowadays.
I've got a Qwen 3.5 running on a 12GB 3060 and it's dumb as a stump but still smart enough to get some useful work done. Since it's my daily driver desktop I havent jumped to 3.6 since last time I did I quickly ran out of vram and locked the desktop environment.
But yeah, the Qwen line is pretty impressive on commodity hardware.
I must be using LLMs very differently than y'all, because I can't think of a single thing I would rely on an LLM that's "dumb as a stump" to do for me.
To me, LLMs are for asking research questions + exploring design spaces + pointing at codebases to investigate bugs. And those all benefit from the model being as "smart" (in terms of both fluid intelligence and burned-in knowledge) as possible.
I'm guessing there exist problems where "intelligence past a certain point" doesn't matter, so these medium-sized models can match the performance of the bigger models. But what problems might those be?
Things that are tedious but simple but I'm unfamiliar with.
"Go add a gh action to compile and deploy this thing and run its tests" is one I've found it's good at. Yes I know how to make a gh pipeline but it's always a hassle to remember what goes where.
Cranking out unit tests is okay. It's good at summarizing things so it's not half bad at writing jsdoc/xmldoc comments.
I also don't get why this twitter user is linked here, versus all the news articles about this new hardware that have been everywhere over the past number of days.
> you may run some models locally if only from a cost perspective
I have a hard time believing running a model on a laptop will be cheaper than running it in a datacenter. Why wouldn't economies of scale apply here as with every other computation?
The datacenter setting has huge economies of scale for low-latency, just-in-time inference using extremely large models, but that's not the only viable use of AI. Batched, unattended inference of possibly smaller and weaker models, while theoretically viable in a datacenter setting, is far from the best use of that hardware. This is where local AI is at its best.
A laptop is really a pretty bad form factor to run LLMs. Worst cooling, more expensive memory that you cannot replace, resell value depreciating fast. It’s fine for tinkering, small scale research, and demos but it’s definitely niche.
The vision NVIDIA is selling is pure marketing IMHO
This is assuming that you'll be priced the fraction of computing that you consumed. But you are actually paying for their infrastructure, for the R&D (and also the computation that went into training the model) etc.
It is not clear that, for your own small computations, this kind of costs are needed, but you will still pay your share in the investment the provider made so that they could serve everyone's computation needs.
But, currently ... you're not. AI companies are operating at a loss, and are being subsidized by their investors.
Local may or may not be cheaper than remote now, depending on the details, but the factors you describe won't affect the math nearly as much as they will once that subsidization ends.
In that analogy bigtech AI is currently investing in cleaner air for all of us? We _could_ breath it through their hose, but might as well breath it outside.
What "every other computation"? I seem to have a lot processing power at my disposal here, between my cell phones, laptops, gaming PCs, various other hardware devices.
You're going to need to analyze the problem much more deeply because it sound like the standards you are implicitly applying would result in "economically, everything should be centrally hosted" but that is clearly not the result that obtains. Even a modern mid-grade cell phone is no slouch; you may not be running a current-gen frontier AI on it but you certainly can do a lot of other rather intense things locally that would have been laughable 10 years ago, like suprisingly high powered games.
AI models will pretty undeniably affect your electricity bill; yes you already own the computer, but it will cost more to run it if it's doing inference!
To a point, but we're talking a laptop, not a server farm. Even if you're going fullbore wide open 24/7 that's about $150/yr in electricity bills at average rates. Not quite nothing but in terms of AI costs that's pretty close to rounding to zero.
I suspect personal privacy and need to run AI workflows to handle the litany of administration tasks of a household will be what result in regular need for local AI.
Apple is already out front with this on a personal, individual level, but they are not obviously headed toward multiuser/family-level ~biz admin with a persistent server running local LLM.
The security aspect is the main driver why I’m seeing so many businesses investing in local hardware. They know the models aren’t as good (caveat that they also can’t run Chinese models) and that’s ok. Places that really care about security and data governance already aren’t on the bleeding edge. They wait for the nice stable lts version, they lock down dev machines in frustrating ways and have lots of IT admin layers.
But they also want to taste the sweet fruit of AI so the only way to do this that a CISO will approve is on local air gapped hardware. It’s a niche but still a billion dollar niche.
You must be unaware that System76 was already selling 192GB machines, mac studios used to be 512GB max. The only reason why we don’t have them anymore is that we are in RAM shortage.
Those 192GB aren't unified memory though. 128GB on Mac or 395 can be used by both CPU and GPU. It's the GPU + large memory that opens up fast local LLM inteference.
Yes, true. But if we had the ability to buy that much RAM in the laptop, everyone would be looking in that direction. Until this thing discussed here comes to the market, “we didn’t have computers with unified 128GB RAM either” (except of macs).
You assume I use the subscription. There are other options but they require more than 128GB unified RAM. You also assume a lot about how I work. And those final assumptions about what and how I think of others speak more about your anxieties rather than what I think.
You assume a lot. Sometimes it’s good to simply ask a question.
Lots of people are already running AI locally. They are the people buying up all the consumer-grade nvidea gpus. What are they doing with them? Well, the same things people with home media or email servers are doing: stuff they dont want to share with the general public.
I want to reduce my dependency on companies like Google, OpenAI, and Anthropic. Aside from the concerns of data sharing I'm also not a fan of how they run their operations, for example Anthropic now using xAI's Colossus data center which is poisoning a marginalized community, or OpenAI getting in bed with the military.
Not everything I want to use an LLM for requires "PhD level intelligence", and increasingly I'm finding more uses that involve sharing my personal data.
Yesterday my local model helped me when looking for a doctor who is in-network for my insurance. I threw it a screenshot from the providers search results and it looked up reviews for all of them.
My local AI is currently upscaling an old british comedy from sub-dvd quality to 1k. (It is not availible other than on DVD.) It looks like it will take about a week for my pair of 5060s to chew through the task.
I own the DVDs so I'm OK upscaling/editing my own copies for my own use. But if I ran the task on an ai service I would no doubt trigger copyright issues.
I hope a family-level AI appliance is a thing later. Local non-cloud assistant that lives in the house, families interact via voice or phones or whatever. Knows the contextual family stuff you need, etc.
We didn't get people buying family-level file servers for the family photo gallery and documents at any real scale, so i doubt we'll see similar for AI especially when the cost is that much higher for GPUs vs an SBC machine.
because nas hardware and software suck and everything else was a poorly executed subscription product...i think one was called helm, another was by early twitter alumni. imagine a home device that manages and maintains itself and is a joy to interact with.
not automatically, but a meaningful step up in ease of use (managing photo/video backup from all family devices) without a subscription would be a solid foundation
> "Ranked in the top 2% of scientists globally (Stanford/Elsevier 2025) and among GitHub's top 1000 developers" - side note but this guy puts this everywhere, gives me probably the inverse of what he is marketing for.
Lol yeah seriously, that stinks "I ask AI to generate a huge amount of bullshit and upload it to pad irrelevant stats".
I agree that it sends the wrong symbol, but actually Daniel is great. He cares tremendously about doing work that is actually real-world useful. I've co-written a few papers with him, and he's really hard working and open to outside suggestions. The danger is that if you send him comments, he'll eventually manage to rope you into writing a new and improved version. Seriously, if you are a non-academic computer scientist with a good idea that you want to publish, he'd be incredibly open to working with you.
As to why he now has this on his blog? I also cringe when I read it. I presume someone told him he should self-promote more, and this is his lame attempt to do so. He's almost certainly the most cited person in his department, but it's entirely possible that none of his colleagues actually know this. Cut him some slack. Self-promotion is not his strength. He's a nerd's nerd, and not a marketer. I'll mention to him that his attempt here might be backfiring when I'm next in contact with him.
I kind of get it in the sense that every academic has to make themselves somewhat comfortable with self-promotion even if they don't like it. It's an important part of getting funding, but putting a blurb like that everywhere just hurts his credibility I think.
He's not a loser; he's done some really fun work that many people use daily. I've used his range mapping trick in multiple projects/papers. It's elegant.
It sounds like he's gotten bad advise about how to market himself /or/ this is being marketed to people who have bigger checks to write and whom he believes will be responsive to this kind of marketing. As an academic, it rubs me very wrong - I think it's detrimental to the field when we get into h-index stacking contests or citation count comparisons. But I don't know what incentives he's responding to, which seems important for putting this stuff in context.
(as an aside, it turns out that polars + fastexcel is about 10x faster than pandas + openpyxl for searching that dataset, if anyone else is curious what he was actually talking about. :)
I think the local-model use case is going to become less niche pretty quickly if the models keep getting smaller and more capable. Even if most people do not care about privacy or offline use, the cost argument is pretty strong
This feels fluff to me on the part of the author (whose work I don’t want to trivialize) but I don’t think they’ve actually looked deeper than a paper spec sheet on this.
1. Yes it has the same number of cores as a 5070 mobile. It’s also running at a shared peak of 2/3 the bandwidth and a shared peak of 2/3 the TDP. The GPU by itself will likely perform at half the dedicated units performance
2. Apple may not have SVE2 but they do have the AMX (private) and SME. I don’t see why he thinks the SVE2 will give him more performance than the SME.
3. He mentions a single core type but doesn’t mention the total makeup. We already have known for a year how the DGX Spark compares to Apple chips. For CPU it’s roughly equivalent to an M3 Pro and for GPU compute (not rasterization) it’s between an M4 Pro and M4 Max without considering bandwidth.
The real advantage to these is that they run CUDA. That’s it. Otherwise when they launch they’ll be 2-3 generations behind where Apple is and 1 gen behind AMD.
The other super power of the DGX Spark was the NIC for pairing them together. But that’s been removed here too.
Prefill is another advantage vs. Apple. It's way way way way faster on a spark than it is even on an m5 max.
Same model, same quant, same query, as close to as matched settings as I can get from vllm, and for workloads with large prompts + low cacheability, one of my sparks will often be done responding before the mbp is done with prefill.
> GPU compute (not rasterization) it’s between an M4 Pro and M4 Max without considering bandwidth
You are likely thinking about token generation which is dependent on memory bandwidth where Apple has an edge. Spark's GPU compute is way higher than even M5 Max (17 FP32 TFlops), around 2x FP32 TFlops... It's literally 6144 CUDA cores like desktop 5070, slowed down by slow memory and lower TDP (29.7 vs 31 FP32 TFlops on 5070).
That’s only if you consider FP32 specifically. On average the M5 Max will pull ahead for tasks like GPU raytracing (it’s currently the fastest mobile GPU for Blender rendering) and token generation and other things that benefit from the higher memory bandwidth.
I’d also mention that you’re comparing peaks which the RTX Spark won’t be hitting. The top TDP is less than that of the DGX Spark.
I just think anyone calling this a beast and a game changer are conflating/extrapolating from different form factors and constraints
It is absolutely fluff, and the only reason this worthless tweet is on the front page of HN is that this audience has a habit of canonizing certain people, and treating each of their bowel movements as prophetic.
Guy suddenly became aware of a chip that the rest of the industry long knew about, seems completely unaware of the competitors, and posts about how it's a BEAST and will be a GAME CHANGER.
Like the DGX Spark was a game changer? Eh, it has mostly been a massive disappointment. An overpriced nvidia laptop isn't going to change the equation an iota.
The Qualcomm Snapdragon X2 Elite Extreme trounces Nvidia's chip in single core CPU performance. It beats Intel and AMD's best, too. It has unified memory. It's the only CPU in the same league as Apple's M-series in both CPU performance and power efficiency. And it's available in laptops today, not later this year. People are sleeping on Qualcomm.
Garbage operating system support. If you can’t do Linux support it’s a bit pointless because there’s two platforms for this that matter: Linux and Darwin.
Qualcomm is like AMD was for GPUs for like decades. Lots of announcements and people on the Internet are huge fans based on web pages they’ve read but if you try to make it work it’s a nightmare.
Snapdragon X Elite doesn’t work on Linux so it’s a pointless platform. Enthusiasts have made M1 work better. Literally have old Macs running rather than use Qualcomm.
It trounces ARM's old CPU design. The X925 used in this Nvidia chip is 2 years old. X930 or C1 has shipped with Mediatek Dimensity 9500 which is what the Snapdragon 8 Elite Gen 5 / X2 Elite should be compared to. Although Qualcomm still has a lead in performance, but it is increasingly shrinking.
But perhaps more importantly. Nvidia seems to be doing a lot better with its ecosystem. Nvidia has much better distribution channels and partners building on top of their PC Gaming GPU. It also have gaming developers relations that is unmatched by any in the industry.
Qualcomm has so far failed to execute this, both in PC and on there Server CPU side.
I'm not sure they are sleeping. I have an older version and it can run games and other things just fine, its just over priced and not properly cooled. The driver/firmware support from Lenovo / Qualcomm is purely garbage. You're lucky to get a driver update to fix anything. For months it just overheated and video would start corrupting but that got fixed finally. You cant just go to Qualcomm's website and download new drivers even though it looks like you can - they really dont get how modern GPU's work on Windows - a driver updates to optimize for games is really something important because of how Windows is but the experience is pulling teeth. If the systems were Neo priced (500-700 USD) and had a cooling fan I'd be all on board with these systems. Right now, AMD with unified memory is just the better deal for the $1200 (2025) systems to run Windows and an average workload.
Seems like not? Judging based on https://github.com/qualcomm-linuxsomething is happening, although I can't say how much. They definitively seem awake at least.
The problem with these chips on Linux is that something has been happening for months but you still end up needing to download special editions of ARM Linux images to get these devices to work properly.
Some distros still need extracting Qualcomm firmware from Windows to get Linux to work properly. Audio remains a challenge, like x86 Linux decades ago. Apparently camera stuff works these days but produces images of subpar quality.
These issues also occur on normal Linux. My experience with my Lenovo+Intel laptop was that it took three months after release for the firmware to work properly (and the Nvidia drivers took much longer, but that's my fault for buying something containing Nvidia hardware). Intel managed to do what Qualcomm did in months rather than years.
I hope Qualcomm finally sorts this shit out, I really do, but with the prices of computers these days, I'm going to need to see quite the discount before I'll consider buying anything with a Snapdragon.
one of the biggest issue i see is the devicetree nonsense. It makes every single laptop and bios version very unique and requires a lot of housekeeping. There are also big chunks of work (as i understand it) to be done around hibernate and decent suspend support.
My experience (wanted to use x13s as daily sriver) is that there was good progress for about a year, until jhovold was leading the charge, but something expired and qualcom as far as i can tell forgot that some progress should happen on x1 and x8c as well as x2.
It feels deeply unfortunate that even with Windows on AArch64 requiring ACPI that it still doesn’t suffice for Linux, unlike on x86.
And I know a lot of that lies on the vendors, but it does feel unfortunate (from a standardisation/conformance/certification point of view) that Windows requiring it doesn’t make it easy to boot other OSes!
They have original thoughts! It's just that those employees get squashed by other divisions or having to meet short term quarterly profits it seems.
There's also the whole giant trillion dollar company doesn't want to invest and let small ideas grow. They only focus on things that move the needle, which isn't much at the size.
Had Microsoft executed and invested, they could have made a come back imo in both search, mobile & hardware. Unfortunately major lack of leadership or they just don't want those areas.
Unless the chip was called Copilot, they are not thinking anything about it. If was called Copilot, they'd have already figured out how to shove it down your throat.
Qualcomm is a “fool me once, shame on you, fool me twice, you don’t fool me twice” kind of situation. So many horrible experiences in the past that people are going to be hesitant.
Qualcomm are trying harder now it seems. But it will take time to repair their reputation in the PC market.
They burned me with the first gen Snapdragon X Elite. Before the various laptops with it were out they promised Linux support. Here we are, years alter, still no fully OOTB support. Ironically, the GPU firmware were just mainlined in the kernel 4 months ago, but they still haven't done the same for the 1st gen X elite.
Tuxedo computers tried and didn't succeed either.
I will never buy Qualcomm again. I avoid them on phones as well by just buying Apple. They do not support their hardware beyond the release.
> I avoid them on phones as well by just buying Apple
To each their own, but I don't recall Apple ever mainlining any of their drivers on Linux. You're rightfully angry on the laptop side of things, but Apple is much worse than Qualcomm when it comes to open source support for their phones.
Qualcomm probably shouldn't have promised Linux support in the first place. Everyone seems to love Apple's hardware even though you're practically stuck with macOS. Had Qualcomm just stuck to Windows-only, they would've probably received a much better reception by the tech press.
Qualcomm has been upstreaming Linux support for some of their chips but they're not working fast enough and I don't think the latest chips are there yet unfortunately.
I've been keeping an eye on the state of Linux on the first gen of X Elite and it's sad that the potential is not fully materialized outside WoA. Take a look at what peeps are going through:
Too bad Qualcomm provides shit drivers for Linux, never updates any of their drivers (had a Samsung/Qualcomm phone with drivers years behind the equivalent Google Pixel phone), etc... They are the absolute worst actor in the entire computing world, don't care how fast their chip is.
Technically speaking, Qualcomm acquired Nuvia, which is where this came from and that company came from ex-Apple engineers wanting to do what Apple said no for their chips.
Why do people care so much about single core performance? We are all professionals here and I bet most of our workloads are multi core. I get that these new arm chips from Apple and Qualcomm are great at one thing at a time, but for professional workloads high end x64 chips still cannot be beaten on the desktop.
What x86 chips have the same or higher number of cores in the form factors that these chips are available in and are also more performant?
Strix Halo is 16 cores. Intel Core Ultra 9 285HX is 24. Apple is 18. Qualcomm is something similar too but I can’t recall. NVIDIA is 20.
Until you get to threadripper/epyc or Xeon territories (completely different form factors and TDPs) the arm chips are ahead on both power and perf than the x86. And even when you get to those areas, arm is equivalent or out performs them as can be seen by the recent neoverse x3 and Vera benchmarks.
Single core performance is the biggest factor for most day-to-day use of a computer, the stuff I do on a laptop. It's more important than peak multi core performance for web browsing and games. I only care about multi core performance when I'm compiling, and I usually do heavy compiles on a remote machine rather than on my laptop.
outside of anything else, amdahls law means that as the parallel performance grows, we become _more_ limited by the inherently serial code, and thus single core performance, not less.
Given that single core performance is "harder" (can't just throw more cores/sockets at the problem), it's also critically important.
I have been somewhat surprised at the lack of commentators observing that this is Microsoft and above all NVIDIA launching a device that is fundamentally at odds with the metered cloud model of AI.
When you look at the other announcements and murmurings (better offline BYOK for Copilot, talk of an unmetered AI future) I think it’s clear that these two firms understand that cloud-only AI is not sustainable or inherently in their interests. But their willingness to undermine OpenAI with a product like this is notable.
I don't think you can interpret this as anything other than a sanctioned rebuke, right? Everyone has a strong visceral sense for what that means.
Copilot just got proper "offline" BYOK support, didn't it? Presumably that was one of the things they were talking about. Though I imagine that has something to do with the fact that Zed has supported that properly for months.
I think most people are not understanding what this kind of laptop will provide.
Before we get local AI, we'll be using hybrid AI.
Running big models locally is unrealistic ($$$$$) but, if you imagine an Agentic Workflow where some bits run on the cloud and other smaller tasks locally, it's an amazing deal. You don't need Opus/Code/DeepSeek/Kimi/etc to do basic stuff that models like Gemma4:12b/Qwen-27b can do locally with much less latency.
Having a laptop where I can use a remote big model and combine it with 5 local domain specific models, is something I would love to do today. Imagine using OpenCode and you've a small model deciding which tasks run locally, then decides if you've a good local model for XYZ task or if we use a cloud model.
My main concern is: Is this hardware powerfull enough to allow local quick models switch? Unlikely but I hope I'm wrong
Given the incredible progress of local models, on present trajectory I think we see comparable levels of performance to frontier models in two years on 128GB unified RAM and 6-bit quantisation. Note how the frontier models are now hitting superior benchmarks with only 200,000 tokens. I think we still have a long way to go with distillation.
> “Our goal is to deliver unmetered intelligence to every home and every desk with Windows,” said Satya Nadella, chairman and CEO of Microsoft. “RTX Spark marks a real breakthrough towards that vision.”
I expect computers with this chip will be about $4000. If Microsoft can deliver on local AI models that can orchestrate Windows and have solid real world intelligence, that will be an inexpensive business purchase compared to pay as you go tokens. I'm excited to see how this plays out.
Enterprise, of course. They probably buy more PCs than the rest of the market combined.
Even for personal use, I'd imagine the amount of people dual booting Windows and something else are a very tiny minority.
Saying "Windows PC" is a pretty reasonable way to distinguish between "made by Apple" and "made by someone else" because the market of PCs that aren't made by Apple and don't come with Windows is really, really tiny.
To be honest, this seems like a strange hill to take such an aggressive stance upon.
> And who in 2026 is still anal-fixated on a "Windows" PC?
I'm assuming it's just clarifying this isn't about Macs.
The term "PC" is ambiguous, since it can either refer to all personal computers in its original meaning, or to the IBM PC lineage that is mainly contrasted with Macs. Remember the famous "I'm a Mac, I'm a PC" ads.
When you just say "PC", people today genuinely don't know which meaning you are referring to. And "IBM PC" is antiquated, and "IBM PC clone" is even worse. So "Windows PC" is a pretty decent name.
Do you have a better suggestion? Because "Non-Mac PC" doesn't exactly roll off the tongue. If you say "Windows PC", everyone knows what you mean.
And it's not an "anal fixation", there's no need to be gratuitously insulting.
And this isn't a "Windows PC" in the traditional sense. The reason people run Windows in the enterprise (and for some desktop home uses like games) is still hardware and software compat.
I run it for work because we make windows programs. We use drivers that don't exist on Win-for-ARM yet. So to most people a "Windows PC" is an x64 Windows PC still. The risk for MS if compat isn't good enough for Windows-Arm64 is that people might as well shift from windows entirely if they need new software and harware anyway.
Hopefully anyone who wants to run anything other than Windows on an Nvidia-produced device has learned their lessons at this point. Although, a cursed Nvidia Hackintosh would be extremely funny.
For normal people, there are three computer operating systems: Windows, Apple, and ChromeOS. Nvidia isn't going with ChromeOS and Apple hates their guts, so Windows is the only normal operating system they can market.
Their marketing makes clear that these devices aren't the piddly Chromebooks that ruined the desktop experience for so many people (expensive Chromebooks were nice, but rare in practice).
Qualcomm promised Linux support, failed to deliver, and now anybody burnt by their promise won't want to buy their hardware again. If they promise a Windows PC, people won't have reason to complain when Linux or FreeBSD or SerenityOS won't boot on there. Given Qualcomm's failures here, Nvidia is probably doing the right thing.
> Although, a cursed Nvidia Hackintosh would be extremely funny.
I did this for years. We ran Resolve color correction suites with external chassis to place multiple Nvidia GPUs in it at a fraction of the cost of the shitty TrashCanMac that was available. Lots of people continued to use the 2012 Cheese Grater MacPro with its older CPUs. The only way to get modern (at the time) compute in a Mac was to use a Hackintosh. Since it wasn't for personal use, not having things like AppStore, Messages, Music, etc wasn't a big deal, so building a Hackintosh was easier.
I built one for personal prosumer use around the time of the 1080s that allowed me more machine for the dollar than Apple offered. Once the M-series chips came out and they were capable of what the Hackintosh was doing for me put me off of building anything newer.
Windows is dying a death by a thousand small, user-unfriendly decisions. This is genuinely sad because the technology underlying Windows is actually very robust and flexible.
So, the partnership is maybe natural, but not prospective. Also, note how Linux is getting popular among gamers. Of course, it's way behind Windows, but the direction of the change is clear.
I'm convinced that Nvidia is not primarily targeting the consumer market and that the ultimate goal for its CPUs is the server space. The company invests effort where the money is, and consumer products account for only a fraction of its total revenue. Maintaining a presence in the consumer market seems more like a way to avoid a complete pivot than a strategic priority.
I follow Daniel Lemire and like his contributions, I also understand that the HN thread was created for discussion purposes, but I'd really appreciate having a reference to the spec or a source to the claims made, either here on HN or on the tweet itself.
I dislike the cycle of propagating news and assuming that someone else double-checked it.
- NVIDIA RTX Spark powers the world’s first Windows PCs purpose-built for personal agents, featuring 1 petaflop of AI performance, industry-leading power efficiency, full-stack NVIDIA AI and graphics technology, and up to 128GB of unified memory.
- NVIDIA and Microsoft collaborate to deliver a native Windows experience for personal agents, including new security primitives and NVIDIA OpenShell to run agents securely on primary devices.
- RTX Spark lets creators, AI developers and gamers render ultralarge 90GB+ 3D scenes, edit 12K 4:2:2 video, generate 4K AI videos, run 120B-parameter LLMs with up to 1 million tokens context using agents locally, and play AAA games at 1440p and over 100 frames per second.
- Adobe is rearchitecting Photoshop and Premiere from the ground up for RTX Spark to deliver 2x faster AI and graphics performance.
- RTX Spark-powered slim Windows laptops with all-day battery life and premium displays, as well as compact desktop PCs available this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, with models from Acer and GIGABYTE to follow.”
I do not see how it is a 'beast of a' anything. It has 300GB/s memory bandwidth, barely above AMD Strix halo (256GB/s) with the same 128GB RAM and less than half memory bandwidth of M5 Max 128GB (614GB/s). Emphasizing memory bandwidth because most people interested in it I suppose are AI enthusiasts. Also, Windows.
They have a lot of software groundwork ahead of them to make an ARM CPU viable for any kind of desktop use outside direct inference or training usage too.
AMD has the advantage that their x86 machines run everything, Apple maintains the whole MacOS stack, while Nvidia can barely scrape together one Ubuntu release per Jetson generation, it's beyond embarrassing. Maybe they ought to put those agents they keep droning about to some actual work on their OS support.
The interesting part to me isn't really the Cortex-X925 vs AVX-512 comparison, but Nvidia trying to make the GPU the center of a Windows PC rather than an add-in card
>but Nvidia trying to make the GPU the center of a Windows PC rather than an add-in card
over the last decade, many software (especially the popular and industry standard ones) shifted to GPU accelerated design. it's a push before NVIDIA even tried to capitalize on that.
guaranteed it will be like Qualcomm arm, it's a partnership with Microsoft after all. we may see a community project to make Linux work on it but it will not have an official first class support and many things will likely not work properly.
Is it really unified memory? AMD Strix Halo is "unified" but you still have to allocate memory separately for cpu vs gpu. Apple Silicon is true unified memory.
My understanding is that this is the limitation from Windows not from AMD SoC. There are several internet resources to "enable unified memory support" on linux eg [1].
As a side note, qualcomm chip set on Android has been doing this for years (like Apple) so it's not super unique thing. It's more like there was no need before.
Even then the "reserved" section is a carve out guaranteed chunk to allow stuff that might need contiguous physical memory (display scan out buffers and page tables, for example) and similar.
The GPU can still happily use all the rest of the memory for other use cases - which tend to be the bulk of allocations anyway. Though there might be performance implications - for example "moving" buffer ownership to the GPU would need to evict CPU caches, and often 4k pages and tlb lookups can be a pretty inefficient situation for GPU-style accesses.
That's been pretty standard for any SoC for decades. And "differences" to apple's SoC are more implementation details.
yes, but more due to OS limitations than hardware. You can use their GTT which is then _true_ UMA where GPU can grab whatever it wants from the memory pool.
This isn't the first time we have UMA on the PC, btw. When SGI did their PC workstations, their 320 and 540 PC workstations had what they called Cobalt graphics chipset and crossbar with their IVC architecture. They bypassed AGP at the time completely. It was quite unique to see strict UMA on a PC. Haven't seen it since until these new systems we're seeing now on PCs and Mac.
For local models, the useful part is not just having 128GB attached to the package. It is whether the GPU can practically use that memory without the usual VRAM-style constraints
> you still have to allocate memory separately for cpu vs gpu
That's an API issue not a hardware issue. Regardless, I believe the major APIs permit seamlessly sharing pointers at this point? (I have no experience doing that though.)
It is unified in the sense that the OS can dynamically assign memory to CPU and GPU. Apple silicon is not a alien tech that other silicon vendors cannot implement.
As he likes to share often, "He ranks among the top 2% of scientists globally (Stanford/Elsevier 2025) and is one of GitHub's top 1000 most followed developers. "
Still, Microslop has repeatedly proven their ability to slow everything down to a crawl no matter how powerful the hardware. If you want it to be fast, don’t use Windows.
Just give it a year or two and Windows will drag that sucker into the mud and run everything just a sluggish as ever.
The idea that any hardware performance increase will be eaten up by terrible software is an evergreen. A computer that could serve as the single server for a medium size enterprise 20 years ago, is no longer able to serve as a desktop for a receptionist. I'm not even sure we're talking diminishing returns anymore, we're probably past the point of maximum yield and into the negative returns at this point.
The dgx spark is the same chip and those are in the low 3s to 5 range for most of them depending on manu, storage config, etc. The dgx sparks also have connectx 7 cards in them to support the 200gbps networking for RoCE.
So I would expect the mini PCs to come in less than the sparks. Laptops I assume will be close in price with the addition of all the other laptop stuff.
See [1]. There's not 128GB of on-chip memory. "Integrated" memory in this context means that the GPUs and CPUs all use the same memory. There are on-chip caches, of course.
It is all in integrated into one monolith “superchip” package. The 128gb of RAM isn’t going to be purchased separately or be upgradable. At least according to all indicators. Which is what I was responding to.
A RTX Pro 6000 has ~24K 5th generation tensor cores, I'm guessing this would then be 1/4 of the count but 6th generation? Wasn't clear from the images.
What is more important than core count is how the caching architecture is laid out. They could lay out those 6k cuda cores in a layout which provides much larger blocks of cache to smaller number of cores. That would increase the memory bandwidth which would be better for inference.
> The memory is not as fast as dedicated GPU memory, but it is cheap enough while delivering enough bandwidth to run AI models locally.
Also "cheap while delivering enough" certainly sounds like someone is trying to temper expectations. It sounds like something sitting in-between GPU+VRAM inference and CPU+RAM one, not as a step above/besides GPU+VRAM.
Having slower memory may not actually lead to lower memory bandwidth. The cuda cores can be broken up into compute complexes which larger blocks of memory directly attached to the cores. These could be filled with read operations from the bulk system memory. You can start executing and then page the next batch of data in while compute is working. For LLMs you don't have much random memory access, you can sequence your accesses in blocks.
If these chips become popular I am sure you will see LLM architectures taking advantage of the parallelism.
Intel's basic architecture keeps accelerators away from main system memory, unlike, for example, IBM's POWER architecture where the CPU and GPU are equal 'users' of memory. It's not a great breakthrough to suggest something different. The problem is - it's different, and not compatible with a lot, or most, or all, existing hardware. Also, there are some security concerns, as @stego-tech noted.
Says running local llms isn’t relevant. Than says it is decent for games, which is just correct if you compare any gpu remotely similarly priced. I don’t understand what is the point he is making
They are useless if RAM prices are this high. $800 laptops with maximum 8GB are currently the norm, Windows 11 can't run on them decently. No matter how fast the SoC is with overpriced RAM they are slow. Systems that can make good use of them with 64-128GB are not affordable anymore thanks to Nvidia and co. This is a smokescreen. They'll probably sell them packaged as compute modules anyway.
while unified memory may offer better performance than unsoldered DDR system memory, it still won't be as great as 1.8TB/s bandwidth on high end consumer GPUs right now.
nvidias master plan may be making it the new normal to have "only" 400GB/s bandwidth, thus gatekeeping local model usage further behind "more memory but not as fast as the cloud can do it"
I think it’s an interesting theory but a bit too conspiracy theory-ish.
Nvidia just wants to sell stuff to everyone.
And I think for professionals doing local AI work, products like Strix Halo and Apple Silicon are a competitive threat.
A big part of maintaining the leading software ecosystem is ensuring you have competitive hardware for all your users.
I also think the RTX Spark product is relatively low effort for Nvidia. Grab a Mediatek CPU and slap an Nvidia GPU on the die. Sure, that’s oversimplifying it, but still.
The competitor for this NVIDIA CPU will not be the now old AMD Strix Halo, but its successor (launched recently), which supports up to 192 GB of unified memory. Thus 128 GB is no longer SOTA.
While this NVIDIA system is inferior from the point of view of the memory capacity, its main advantage is that the top models will have a bigger GPU, i.e. with 6144 or 5120 FP32 execution units, compared to 2560 for the AMD GPU (compared to the NVIDIA CPU, the AMD CPU has a better multi-threaded performance for legacy programs, and a much better multi-threaded performance for the applications that use AVX-512).
However, these top models with big GPUs will also be much more expensive than the competing AMD system, while also being much more expensive than a laptop or mini-PC with an equivalent discrete NVIDIA GPU (which has the disadvantage of having direct access only to a much smaller, even if faster, memory).
I don’t think there is much improvement in compute for the new strix halo revision. The next one supposedly adds rdna4 cores or similar and more memory channels
I have a 128 GB LPDDR5X machine. It's a great workstation laptop (which is why I got it) but the memory bandwidth is just awful if you're wanting to use it for AI. An old Epyc CPU will fair better both in terms of being able to run full sized larger models as well as having higher memory bandwidth, and that's not a recommendation to go that route either as it's still not worth it.
Qualcomm is too. They mainlined the GPU firmware for the X Elite 2nd gen, but still have not done so for their 1st gen X Elite which they promised full Linux support for and failed to deliver, and have now moved on pretending they never said that.
I’m not sure if you’re aware but there is a supply chain shortage for pretty much everything needed for a PC that isn’t expected to be solved this year or next year. There is no way that can be affordable
or moreso, the available supply has been eaten by rampant speculation, and hyperscalers have overpurchased vs the datacenters they can actually get built and power
Don't want to be too harsh, maybe I'm missing something, but the CPU is at least 2 years old, internally it has been a complete shitshow and that's a minor hiccup when compared to the firmware and software situation.
It's an interesting "newcomer" and the more the better but calling this a "beast" and a "game changer" is ridiculous to say the least.
> The game changer is the unified 128 GB memory. That is the path Apple took years ago. Instead of separate memory for the CPU and GPU, everything shares a single pool. It is increasingly popular.
> The memory is not as fast as dedicated GPU memory, but it is cheap enough while delivering enough bandwidth to run AI models locally.
So, the reason "dedicated GPU memory" is fast, isn't because it's "dedicated"; it's because the types of memory built into GPU cards — GDDR and HBM — are designed for throughput over latency.
Which is to say, GDDR and HBM memory could be shared with the CPU in UMA while still being "fast" (for GPU use-cases.) In fact, the PS4/5 and Xbox 360 / One X / Series consoles have UMA architectures that use GDDR memory as their main memory, with no regular DDR memory to be found.
What I don't understand: why don't we see UMA architectures where there's both regular DDR and GDDR/HBM memory mapped into the address space of the CPU+GPU? That seems like the best of both worlds: you'd have some memory that's "tuned" for random-access CPU usage (regular DDR), and some memory that's "tuned" for streaming GPU usage (GDDR/HBM), but either type of memory can still be put to the use it wasn't "tuned" for, just with slightly-worse performance.
I guess you'd need to do a bit of software work:
1. a bit of work in the OS kernel / malloc library to get CPU workloads to "prefer" allocating DDR memory over the GDDR/HBM memory until they've exhausted DDR memory (or maybe not, if you just tell the kernel the GDDR/HBM memory is something like a zswap thinpool);
2. and a bit of work in supported ML frameworks, to teach them about a hybrid strategy between UMA "allocate anywhere, it's all the same" and NUMA "keep assets in VRAM if possible; if you spill assets to RAM, then they must stream into VRAM on access" (i.e. "at allocation time, allocate as if the system were NUMA, VRAM first then spilling to RAM; but at execution time, use the UMA codepaths, no need to copy RAM into VRAM.")
Laptops shipping with less RAM is exactly the reason to be interested in native apps again. Every app being a chrome/EdgeWebView process is the problem.
Is this somehow satire? This is just the dgx spark with keyboard and monitor in a convenient format. Since it has more stuff, I'm sure that the price mark up will increase too.
Up to $5000 because why not?
With that money you can build a real PC with rtx 5090!
The obvious comparison here is the M5 Max where you can buy a Macbook Pro with 128GB of also unified memory. Obviously CUDA cores are specific to NVidia so it's hard to directly compare but I've seen claims that the M5 Max is roughly equivalent to ~4000 CUDA cores. This obviously depends on workload and whether the CPU supports the precision you want to use (eg FP4).
The M5 Max has memory bandwidth of 819GB/s. The RTX Spark I believe is ~600. So it might be slightly better than the current generation of Macs but likely worse than the expected M5 Ultras of the new Mac Studios (likely Q3 2026).
For comparison, a 5090 has >20k CUDA cores and 1800GB/s memory bandwidth with 32GB of VRAM. The RTX 6000 Pro (at ~$10k) has 96GB of VRAM, same bandwidth and ~24k CUDA cores.
We have to see what RTX Spark systems sell for but the DGX Spark is in the Mac Studio price range (~$4k).
I do think Apple has a real opportunity here but there offerings aren't quite there yet. The M5 Ultras might be a really attractive option for local LLMs. I expect them to be in high demand.
> I've seen claims that the M5 Max is roughly equivalent to ~4000 CUDA cores
Who claimed that? The M5 is still a raster focused GPU, dedicated matmul blocks be damned. For some workloads that napkin math might work out, but for many others it's a wild overshoot. Time-to-first-token still favors CUDA, and real-world training workloads aren't getting anywhere near Apple Silicon.
All of the memory bandwidth in the world is useless if you spend 15 minutes processing 64k tokens worth of context prefill. This is where CUDA shines.
Does this person know that this is the same GB chip in the DGX Spark? It isn't some proposed thing, it's a chip loads of people have on their desk right now, and there are endless benchmarks of it.
Decent single core (a long ways from Apple level, but decent), but it makes up for it in cores to provide M5 level performance, CPU wise. Memory bandwidth it is kind of starved, at 1/6th many GPUs.
They got Microsoft to customize Windows for the RTX Spark, and will likely have to brutally throttle it when running as a laptop (it's literally a 140W TDP chip), and that's neat. It's going to be a very expensive laptop.
This is probably the better way to frame it: not "Nvidia is proposing a new CPU system" but "Nvidia is trying to move an existing GB/Spark-class platform into a Windows PC form factor"
The 900 GB/s is from the NVLink-C2C interconnect, if you were wondering about that. They quote "up to 900 GB/s of bidirectional bandwidth between GPU and CPU".
Mind you thats not to/from memory, which indeed only has 273 GB/s.
Yes its not an "Apple M killer" at all. Also, the available official performance numbers are partially overstated (1 Petaflop is only possible for sparse FP4 models, "in theory").
Perhaps a sobering rule of thumb: if it was actually useful, you couldn't buy them because someone would scoop them all up to shove them in a DC and make money with it.
It uses LPDDR5X instead of VRAM and will still sell for a premium while pushing their presence even further in every side of the AI market. This was one area AMD was ahead in and now Nvidia is probably better off making this to compete on that front while still being better off than making a 5090.
That doesn't answer the question. If the high margin enterprise GPUs are saturating the fab capacity you wouldn't expect them to be pushing this. But IIRC those all have oodles of integrated HBM at this point so I wonder if fab capacity for that has become a bottleneck.
I believe it does - the reasons why are exactly differences like LPDDR5X vs HBM3e. Not every fab is capable of making any type of chip another fab makes. If you can make a product with different chips and still sell it for a premium why would you not just because the fabs for your DC product's chips are busy?
Looking at it more, I believe the story repeats with the TSMC processes used for the CPU vs chips like GB200 as well.
Even if none of the above were the case, the question still isn't "why not make the enterprise GPU" it's "why not make the higher margin per chip area product". If the NV1/GB10 take less die space and cost a lot it's not immediately apparent the enterprise GPU actually nets Nvidia more $ per die or not. That's why it's relevant these will be sold at a premium.
This is an enterprise offering. It'd take a guess its to try and stop the bleeding over to macOS. This launch, plus WSL containers, their own de-bloat winget config, mxc, etc. all seems like they're saying "pls stop leaving for macOS, see, Windows can be a great dev machine too."
I think it's niche now because getting the hardware to run it is expensive and the quantized models don't work as well. If those improve then it would be a no brainer to pay one off for the hardware instead of a fortune for API calls.
I am not really convinced that four bit quantisation is that bad; almost certainly six will be enough. But Google are making claims for their QAT tech in Gemma that they are surely using or testing in Gemini that it preserves nearly source model quality while reducing footprint.
The hardware for 50 tokens per second with a four bit quantisation of Gemma 4 26B or the sparse Qwen 3.6 is not really that expensive: it’s a secondhand M1 Max.
Beyond that, I agree. I think moving planning tasks to local is a now thing, not that it really has much impact on token spend. I also think many small coding tasks are fully within the grasp of the above two models.
The main issue right now is that the software landscape is rather confusing, but I reckon uncomplicated Gemma 4 26B QAT support with MTP is a few weeks away.
I think it is likely to appeal to video and photo editors who want to use AI tools (the press release has a quote from Blackmagic Design, as well as from Adobe, who I think have no stomach for their own cloud AI).
But I don’t know about specialised: this could run quite large models with MoE.
Performances of local models are pretty bad compared to what AI vendors offer, token generation is just too slow to be that useful. And you need to allocate GBs of memories, something that will stay very expensive to buy for a long time.
Running local models will stay niche for a while, unless we see breakthroughs
That's a fairly obvious idea, not dumb at all, but unfortunately it doesn't seem to pan out. Trying to specialize an LLM in one area harms its 'cognition' in all areas. For instance, if you train a coding model without all the Shakespeare and soap operas and Wikipedia and pirated Stephen King books and ancient Roman history and whatever, you end up with a worse coding model.
Nvidia is milking the market now. We need more competition again - currently we have a mafia control the prices, not just Nvidia but all the AI companies. The price increases should be paid for them, not by us. "Free market" is being manipulated by them here.
All depends. The current technology will be cheaper in a year or two. The best cutting edge stuff will properly be even more expensive. But in 10 years time... we can run current SOTA models (or models that are equally good ) on our local hardware
We had a thing called globalism that drastically reduced costs. Globalism right now is on life support. Given geopolitics, I don’t see how it’s going to survive.
"I am not sure how many people will run AI models locally. It still seems like a niche application to me."
Clip me :). You are currently living through the final stages of unrestricted computing in the hands of the 'public'. Our regimes are going to pull up the drawbridge in the name of 'safety'. Download the open models asap and prepare for an airgapped computing environment. That will be your frontier in not extremely neutered AI in the near future.
The Unified Memory pool is what will continue to be the “game changer” in systems architecture, especially outside of data centers.
The reality is even cutting edge games and consumer workloads don’t actually take full use of the PCIe bandwidth of the GPU or the bandwidth of its GDDR memory. Even local AI use cases don’t substantially or meaningfully benefit from faster memory, at least to average consumers.
A unified memory pool does two things:
1) Lets systems optimize utilization based on need, rather than be confined to specific pools
2) Reduce overall memory cost, by letting system builders purchase a single type of memory in bulk instead of having to figure out GDDR vs DDR memory placement (important for SFF/portable machines)
So at a time when memory is expensive, unified pools make more sense. Even when memory becomes cheap and plentiful again, it’s just practical at this point to allocate a larger overall pool instead of managing discrete sets.
The one big drawback is security. A shared memory pool means side-channel attacks against memory from the GPU or CPU could potentially compromise the other as well, meaning memory-safe designs are going to be critical to security going forward (which is good for Rust adherents, I figure).
> The reality is even cutting edge games and consumer workloads don’t actually take full use of the PCIe bandwidth of the GPU or the bandwidth of its GDDR memory
Game dev here. For anyone reading this - it’s not because we’re lazy, it’s because _it’s really hard to do_.
One of the biggest differences between the current generation consoles and the current gen PCs is unified memory.
How much of that difficulty comes from the chosen game engine? I assume the engine is the primary factor in how resources are allocated.
One related question that you need to follow that with is the associated costs of switching the whole studio to another engine that's technically better, or if proposing teach studio tailor-make their own engine the costs of that engineering, if presumably they have or learn the expertise to surpass whatever they're using currently.
I'm not a game developer, but it would also seem to be a link between resource usage by the engine, and whatever content the production side are making. For all the commentary about how brilliant the id software engines are, if you examine the levels you pass through they're also very efficient with what they demand out of the engine - it's like an orchestra playing well together, not one instrument that means you can do anything.
And conveniently, by making your machine non upgradeable, it allows the manufacturer to enforce market segmentation / charge a huge premium for small RAM upgrade (a la Apple)
It doesn't -have- to be that way necessarily...
LPCAMM2/SOCAMM2 exist, heck I think Framework is using LPCAMM2 in one of their new laptops.
Heck, I'm willing to bet that a lot of manufacturers would rather go that route than soldered in, if for no other reason than the relative cost of warranty work between the two.
However, people probably need to stop being obsessed with ultrathin laptops for that to happen.
Sir! I am typing this on a Lenovo Carbon X1, with soldered on ram, and you are EXACTLY CORRECT!
I would much prefer two SODIMM sockets with the option to go to 32MB shared video memory, or DDR4/DDR5. Give me OPTIONS!
Until LPCAMM2 came along, using low power LPDDR RAM meant soldiering RAM to the motherboard.
If you wanted to get sleep right and improve battery life, that was the trade off.
There is LPCAMM2, if manufacturers want to use it.
So, it does not have to be soldered.
Is that required or just a choice Apple made?
What do you mean by required? Apple's prices are notoriously disconnected from the cost of manufacturing.
I mean is it possible to make unified memory systems with good performance or is it not really feasible due to memory timing/trace length issues?
It’s possible if you’re willing to go with much slower RAM than GPUs like but CPUs often use. Thats what integrated graphics laptops have done for a long time right?
But can you get high end CPU and GPU performance with unified memory and maintain user upgradable memory in a reasonable way? Thats what I don’t know.
> I mean is it possible to make unified memory systems with good performance or is it not really feasible due to memory timing/trace length issues?
LPCAMM and similar solutions exist, but have never been demonstrated running at speeds that match what the leading soldered memory systems are using; there's always been some speed penalty. I'm not sure we've ever seen a system demonstrated using LPCAMM or similar for a 512-bit bus to match Apple's Max tier SoCs, so it's somewhat of an open question whether those solutions can offer upgradability at the high end of the market for unified memory systems.
You mean Apple prices are notoriously over priced, over hyped, under powered, and
"Abdul Jabar, couldn't have made these prices, with a sky hook."
both. soldered ram is faster. also Apple don't want to offer upgradblity after purchase.
Don't I/you wish. The mechanical junction adds no delay, only manufacturing expense, and the delay of purchasing new systems to keep up with OS bloat.
Actually the opposite is true. Socketed RAM can be made to overclock and adjust timings, while soldered ram, no. Two Lenovo's one soldered ( Carbon X1 ), one T590, one slot: Crucial 16GB, 260-pin SODIMM, DDR4 PC4-19200. Exact same processor, the X1 is DDR3 soldered on 532.0 MHz PC3-1066. The T590, has DDR4, PC4-19200, 1200Mhz.
Both have a Core i7 8665U... and the T590 is much faster, with socketed ram.
I think you'll find that in the current day, high speed LP(?)DDR5 requires a better signal path than what the SODIMM can provide. Which is why laptop makers initially moved to soldered RAM before moving to CAMM (probably only for the high end ones).
how about the LPCAMM route? Framework uses LPCAMM2 in 13 Pro laptop mainboards and claims that it satisfies the iGPU and NPU hardware without needing soldered RAM
Are there no PCIe standards that are sufficient to support both use cases?
What happened to PCIe 8 and CXL?
AFAIK PCIe6 just started getting implemented in hardware last year... PCIe7 Spec was just released last year too...
PCIe6 is a much larger change than 'just bump up the transfer rate', the encoding changed too (on top of the new code length, it's no longer NRZ,) so everyone needed to design and validate both the new encoding block, negotiation, etc etc.
That said, I'm guessing PCIe7 will be a 'smoother' transition from PCIE6, i.e. we might see 7.0 products in 2027. That will theoretically get you ~240GB/sec, on an x16 link, or hypothetically a little less than the hypothetical max of a current Strix Halo. (I'm guessing however, that PCIe protocol overhead will make the difference larger.)
Don't really buy the economic argument. For 99% pf all workloads you need at least an order of magnitude more system memory than gpu memory.
Most systems barely need more gpu memory than what is required for video, browsing etc.
Just because we found a new usecase doesn't flip that on its head.
Besides, I want to keep doing what I'm doing today. So if I need 128GB today and my local AI needs 128 GB then I'd need 256 GB to keep doing the same work.
The argument rather seems to be that we shouldn't use such expensive memory on the GPU. Which might be true if you only want to do inference on it.
DRAM optimized for CPU usage looks very different from DRAM optimized for GPU usage. You are leaving a lot performance on the table when you have a unified memory architecture. It makes sense in some situations, but it is not a silver bullet.
The "one big drawback" is the lack of consumer upgrades, and the seemingly arbitrary prices charged by vendors for memory upgrades at time of system purchase. I'm not saying it has to be that way, but seems like it has been so far :-(
And here I am with 128GB Strix Halo longingly eyeing the Blackwell cards that spit tokens 10-20x the speed.
The question is ultimate shape of knowledge compression and bandwidth optimization at which we arrive I suppose.
If you haven't already, check/increase the GPU memory carve-out on your UEFI.
More details: https://rocm.docs.amd.com/en/docs-7.2.0/how-to/system-optimi...
That was the main reason for the big hype around Memristors 15 years ago. High density, high speed persistent memory to completely remove the need for hdd/ssds, potentially even removing the need for external memory altogether. So frustrating that it still seems like we're a long ways from that becoming reality. There's some renewed interest in Memristors as they can simulate neural network connections in models, so maybe the funding will return for it.
The one example of persistent memory that managed to reach the mass market was Intel Optane/3dXPoint (still popular today among people looking to save on RAM costs) and that used a kind of phase-change memory, which is but tangentially related to memristors. ReRAM is somewhat closer, but it's also been less successful so far.
Well, back in the day... The MacIIfx had video memory, ( dual ported ram ) that could be read and written to out of different ports. Wicked fast. It 486DX2s more than a year to catch up.
Memory safety is orthogonal to side-channels, and hardware-enforced isolation (e.g. IOMMU) is more powerful than compiler-enforced isolation (but both are good!)
Oh no now I have to worry about shaders row hammering my OS ram /s
You really do have to worry about that!
What is the difference between unified memory and shared memory?
Shared memory existed since the first CPU with an embedded GPU came to market and you could set in BIOS how much memory goes to what component.
I do have an opinion about how unified memory could be different, but I want a proper explanation.
I'm not sure everyone uses the terms consistently, but the difference is that the old "shared" memory was reserving a section to act as VRAM under the control of the GPU, ignored by the OS. The CPU ran the same kind of code pretending there is a "bus transfer" between host memory and graphics memory.
In unified memory, all the memory is host memory and data can go from program to GPU with zero copy movements. The addresses of buffers can be shared via appropriate MMU translation support, so that the application and graphics subsystem are communicating effectively through the basic RAM cache coherency protocols over the same buffers.
Edit to add: Aside from the zero copy transfer potential, it also means dynamic allocation strategies can shift the balance between host and graphics allocations on the fly. Individual image and message buffers can be allocated on the fly instead of setting a static split between the two worlds.
Reserved sounds like it would have been a better term now that I'm reading this many years later.
That's my understanding, or, maybe a better word would be "guess". The CPU telling the GPU: this is your memory now.
For these in specific, they appear basically transparently to the GPU. There's a lot of software/firmware stuff for this, but also a different hardware architecture - while the RAM is on the CPU die, the nvlink-c2c gives it extremely low latency and 600GB/s bandwidth between the GPU and CPU.
Shared memory of the past meant reserving a part of the memory for the GPU, which could then not be used or accessed by the CPU. If the CPU wanted to access something, it had to copy it from the GPU's section of the memory to its own. Unified memory means both just fully share the same memory.
Marketing, mostly? But perhaps also more flexibility with how much memory the GPU can directly access without reserving it.
No. Let’s define terms, as others have pointed out they’re not perfect.
Unified memory is what Apple is doing, other phones do, and many low end built in GPUs have done in PCs for ages. There is only one physical memory pool. Both the CPU and GPU can access it at full speed.
This means no copying between pools of memory. No speed penalty accessing the CPU memory from GPU or vice versa. If the GPU only needs 2 GB to draw the desktop it only uses 2 GB of the pool. Or it can use 45 GB if it needs it and the CPU doesn’t. But all memory has to be the same speed, and that ain’t cheap given how fast GPUs like things. I don’t know if expandable memory is possible, and they use the same bus do they compete for bandwidth. Seems theoretically easier to program for to me.
The opposite is what’s been common in graphics cards since the 2D era. CPU and GPU have their own memory and can talk over PCI/AGP/PCI-E. This is what I think they mean by shared memory, if it’s not what’s the point in touting unified?
In this model if the GPU uses 2 GB of its 12 GB total, the other 10 isn’t available to the OS at full speed and I’m not aware of any operating systems that would use it for programs/cache by default. If the GPU needs 45 GB… too bad. You have to page things in and out of GPU memory over the much slower system bus. Starting a game means loading assets into main memory then transferring them to the GPU (newer tech can accelerate this). But the CPU can have slower memory than the GPU saving money. Memory expansion on the CPU side easy. And the CPU saturating its memory bus has no effect on the speed of the GPU memory bus because it’s physically separate. More complicated memory model but it’s the one everyone uses used to.
Which is better is a matter of opinion and workload needs.
Yes, I know there is an actual difference vs. dedicated GPUs with their own VRAM. I say it's marketing because Apple popularized the unified memory term even though, as you said, it existed in iGPUs long before Apple Silicon and was called shared GPU memory.
> I don’t know if expandable memory is possible
It technically is. These new systems (mostly) get their high bandwidth by using more channels (wider bus) of normal RAM modules. A system that has LPCAMM2 sockets should allow using the same LPDDR5X memory but you'd need a socket per two channels. A typical PC only supports two channels so having four (two sockets) would double the bandwidth.
Bandwidth by going wider, not faster. That makes sense.
System RAM has much lower bandwidth and less predictable access. Notably, the transfer from system to GPU is very slow. About 30x slower. LLMs aren’t designed to queue or parallelise operations to account for this. They just become much slower.
Unified memory is only a feature because NVidia so aggressively uses VRAM for market segmentation.
The 5090 ($2k MSRP but realistically $3-3.5k) is almost the same as the RTX 6000 Pro (~$10k). Same memory bandwidth (1800GB/s). Slightly different CUDA cores (21k vs 24k). Big difference? VRAM (32GB vs 96GB).
NVidia ultimately doesn't want to upset this segmentation so the RTX Spark will never undermine their other offerings. This is why I think Apple has a real market opportunity if they choose to embrace it.
To this day I do not get why Intel doesn't just offer massive memory options for their cards. Just charge what it costs to add the extra memory, no upcharge, and they will never be able to keep up with demand. Cheap VRAM is enough to justify a lot of open source investment into challenging CUDA.
> To this day I do not get why Intel doesn't just offer massive memory options for their cards.
They seem to? Intel Arc is the cheapest option by far for a discrete card with 32GB VRAM.
They took longer than everyone expected and then shortly after release they made announcements that made people worry that Intel might kill the project the way they tend to kill GPU projects.
(I still kinda want to get one tho.)
That’s not massive, though. Make it 96GB at $2,000 (ok, probably impossible right now, but they could have before the surge in prices) and you’ll see developers work really hard to make AI tooling work for their cards, CUDA be damned. The same goes for AMD.
It’s like they both want to rely on market segmentation for VRAM too but fail to realize that it’s their only potential inroad right now.
If you buy three 32GB GPUs, that's 96GB total at a very reasonable price. An AI model splits easily by layers, so running on multiple GPUs is quite feasible.
Missed a zero here.
Needs 320 GB Vram
Memory is just one part. AMD has had offerings competitive to NVIDIA for quite some time, but nobody uses AMD cards.
The biggest advantage with NVIDIA is CUDA.
Even low-VRAM cards are actually very useful for running the comparatively smaller dense layers in large local MoE models. This only requires transfering very small amounts of data across the PCIe bus (similar to pipeline parallelism) so it fits nicely around the existing bottlenecks on that hardware.
I have so many questions… Since Apple already sells unified memory systems, what is the market opportunity you envision? Do you see Nvidia and Apple as competitors, and how? (And I’m not suggesting they’re not, necessarily, but I want to hear where you’re coming from, and they do have very different markets.) Hasn’t Apple used storage size (RAM & disk) for market segmentation for decades? And how does a machine with 128GB unified mem not potentially cut into some people’s reasons for wanting a 96GB GPU?
I'm not the person you're replying to, but I wholeheartedly agree with them...
Quick background: doing AI inference requires three things. Lots of memory, lots of memory bandwidth, and of course plenty of compute that has access to that memory.
Quick reference: nVidia 5090 has 1,792 GB/sec bandwidth. 3090 gets about 1000 GB/sec. DGX Spark and AMD 395 whatever get about 275 GB/sec.
Apple M1 Max gets 400GB/sec, M5 Max gets 614GB/sec. Ultra variants get 2x that bandwidth, base variants get 1/2 that bandwidth. However... their compute is rather weak.
Right now, Apple's offerings are juuuuuust fast enough to run dense 27B models at usable speeds at like, 10% of the performance/watt of nVidia. They're world-leading general purpose CPUs but not killer GPUs.
By all accounts, these Windows PCs nVidia is touting seem to have DGX Spark like performance, which is less than impressive. Same with the upcoming AMD AI-oriented consumer stuff.
The other context here is that running your own AI at home is just starting to become feasible in terms of open model availability and the ability to run it at usable speeds. Many are interested in it for reasons of privacy, security, and cost certainty vs. buying tokens.
nVidia and AMD can't make their consumer offerings too good at AI, because that risks interfering with their higher-margin data center sales.(And, let's face it. Even if nVidia did release a 6090 with 64-128GB of memory for an affordable price, consumers wouldn't get their hands on them anyway because people would just start filling data centers with them)
So.
Now you see Apple's opportunity, right? No data center sales to interfere with. No relationship with nVidia or AMD to worry about.
They could choose to make an absolute beast of a home AI machine. The M5 Ultra, if announced, might be that. It's admittedly a niche market, but people are already buying 64GB+ Macs faster than Apple can make them and they're fetching high prices on the used market as well.
The only real questions are if this market is even something Apple would find time to care about, and if they could secure enough DRAM to make a go at it. They are enormous obviously but they're feeling the RAM pinch just like everybody.
They use different technology for their VRAM though. Apple, AMD Strix and NVidia DGX/RTX Spark use LPDDR, whereas discrete cards will be either GDDR or HBM. That directly impacts the memory bandwidth figures. As for compute available, Apple and AMD still have very good figures there for what's essentially a general-purpose iGPU that ships as part of the stock system, rather than a special-purpose piece of dedicated hardware.
There’s something else. Memory size.
Even if a Mac isn’t the fastest in raw numbers it may be faster if it can load the whole model in its ram (went up to 512 GB before shortages) than a couple 32 GB cards could with the data having to be constantly loaded over PCI-E. Because unified memory means the Apple GPUs can access all 512 GB at full speed.
My understanding is this is the advantage that’s pushing huge Mac Studio demand. Because it was the only way to give GPUs so much memory at price points anywhere near.
Yeah you can do way better once you’re in the 5 digits. But below that Apple had a specific advantage for some.
Apple offers relatively affordable options for a high-memory workstation that uses unified memory. They previously offered 256/512GB Mac Studios (both discontinued). Because of this they can keep larger models in memory.
BUT you just can't compete with NVidia performance for LLM workloads (mostly inference) for two reasons:
1. The memory bandwidth just can't compete with a 5090 (1800GB/s). The best current Mac is ~900GB/s. That directly caps tokens/sec and might be manageable but there's another problem; and
2. The raw FLOPS just can't compete with even a 5090. It probably needs to natively support FP4/FP8 to at least maintain a number format parity with NVidia. But beside that, NVidia just has more raw FLOPS.
According to Google, an M5 Max does ~70 FP16 TFLOPS while a 5090 does 380. If Apple can close that gap to at least be competitive and also hold larger models in shared VRAM, that would be a competitive advantage and it would directly attack NVidia's market segmentation.
The Mac Studio last came out March last year. So we may get an update in Q3. Many are pinning their hopes on this. But it might not happen until next year. When it was released the M4 was the state of the art and it came with either the M4 Max or M3 Ultra (which, as I understand it, is basically 2 M3s stuck together, kind of). What people are hoping for is an M5 Ultra with >1000GB/s of memory bandwidth, ideally 200+ FP16 TFLOPS and hopefully FP4/FP4 support.
You can chain Mac Studios together into a cluster with TB5 too.
But it's reasonably likely that the next Mac Studio will be only incrementally better than the last generation.
What should Apple do, in your view, to "embrace" it?
> 5090 ($2k MSRP but realistically $3-3.5k)
These days, more like >$4.1K (at least in the US).
If this thing only has as much gpu bandwidth as the spark, it’s kinda pointles
Not true. This is aimed squarely at the Strix Halo and Mac markets. It's basically just strictly better than the Strix, and it's not clear cut vs that Macs in any sort of blanket statement.
My M5 Max 128gb MBP decodes faster than one of my Sparks, but the Spark's prefill is so much faster it can often answer the same query before the mac's prefill is finished. If you have large prompts, low cacheability, etc., a spark might be a very good options.
Not to mention you get can get two sparks and the MBP will be 85%+ of the cost at half the RAM.
I'm kind of tempted to pick one up. Leave running big models to my dual dgx setup, and all the misc. random stuff on an rtx.
Prefill will be a huge deal if batched unattended inference of SOTA models (on consumer platforms) becomes viable, because at that point it's the main remaining bottleneck. If running 30 inferences together boosts your decode throughput to 3x (that's consistent with some very rough experiments, though these haven't even looked at trying to mask SSD offload latency just yet), that's a 10x in total decode time but a 30x in total prefill time, because prefill workloads are fully compute bound already on consumer platforms and don't benefit from batching much at all.
> The Unified Memory pool is what will continue to be the “game changer” in systems architecture, especially outside of data centers.
The ps4 was the prime example of this, and how it could run so many great games.
yeah, you only see double digits in performance degradation from going from pcie 5 to 3 with a 5090 (at x16 speed), with everything else its like in the single digits area.
And the thing we gamers forget is that we’re the outlier. We’re the edge case.
Most consumers will never really care about, let alone see, the difference in PCIe or memory bandwidth impacts from such a shift to unified memory pools. We might (being, at least in my case, a huge nerd), but I’m increasingly of the opinion that if modern blockbuster games are built for upscaling/reconstruction anyhow, then suddenly such sacrifices to performance seem acceptable relative to the gains in efficiency.
Well I mean, the idea with games is it all fits in vram. You really don't want to be thrashing. It's that things are still so slow that they must be avoided entirely, no?
No copy unified memory will help with that but you do pay the read speed costs.
gen3 is 16 years old.
It’s also the reason, why you will never be able to repair or upgrade your computer in the future. From technological point of view these are indeed big advancements.
However, I couldn’t care less about faster CPU when:
1. It limits my ability to upgrade my system
2. Windows gets increasingly bloated and slower
LPCAMM2
Intel was doing UMA with their i740 graphics in the late 90s. Codename TIMNA was cancelled, but they pioneered it and used it on their you/cpu chips as well as their breakthrough 810 chipset that dominated graphics market for a decade. It was despised because it wa ubiquitous and a low performing graphics engine but games had to accommodate it.
Funny that it is getting credit only now.
> The Unified Memory pool is the “game changer”
M1 knocking from 2020.
Gamed changed, past tense, six years ago. This is catch-up.
Hell, SGI O2s from 1996 had this. For all of the hype the performance gains were pretty modest.
FWIW, the O2's UMA let it handle far more textures than almost any other contemporary system with reasonable performance.
Most other SGIs had single or low double-digit megabytes of texture memory, whereas the O2 could host one gigabyte of unified memory and use a huge chunk of that for textures.
UMA was never about performance and it still isn't. Spark is slower than a 5090.
did they learn why? were there other gains?
I want unified but not uniform - everything can address anything, but you can add slower RAM to the system without requiring an entirely new chip. NUMA is cool.
FYI it existed long before that. Shared memory between CPU and iGPU has been a thing for a long time.
Zero-copy shared memory?
yes, here is 2013 AMD presentation of the topic as example: https://events.csdn.net/AMD/GPUSat%20-%20hUMA_june-public.pd... see slide 14 especially
Ah. Well, what kind of consumer hardware/software combo could I purchase to use this? outside of perhaps the... PS4?
Everything that doesn't have a discrete GPU has unified memory these days. If you're asking for something closer to the RTX Spark or Apple Silicon then look at AMD's Strix Halo systems.
AMD Fusion knocking from 2010.
> (which is good for Rust adherents, I figure).
As a Rust adherent, please do not put words in our mouths or set up unrealistic expectations for other people by linking together concepts at a very shallow level.
Language level memory safety has no answer for hardware security flaws which is what side channel attacks are. No programming language can provide memory privacy if another chip in your machine can read your memory. Just like no programming language can protect your application from a kernel vulnerability of the kernel it’s running on.
Damn. That wasn’t my intention at all, I was just pointing out that Rust has another reason to see wider adoption vis a vis the usual Valley advertising bullshit of deliberately conflating hardware security with software security. I personally give no fucks what something is written in, only that it’s written well enough that I don’t have to twist arms or babysit yet another sloppy piece of code in my enterprise.
But... it's rust.
"I am not sure how many people will run AI models locally. It still seems like a niche application to me. However, it will make decent machines to play video games."
I don't know who will be the winner but with some of the recent releases from gemma it seems more probable that you may run some models locally if only from a cost perspective, not even considering business security. Not sure how this type of architecture would make for good gaming though, puts into question the whole statement.
"Ranked in the top 2% of scientists globally (Stanford/Elsevier 2025) and among GitHub's top 1000 developers" - side note but this guy puts this everywhere, gives me probably the inverse of what he is marketing for.
"I am not sure how many people will run AI models locally. It still seems like a niche application to me. However, it will make decent machines to play video games..."
This is the 2026 edition of Ken Olsen: "There is no reason anyone would want a computer in their home"
> This is the 2026 edition of Ken Olsen: "There is no reason anyone would want a computer in their home"
Digging into this:
> In conclusion, there is evidence that Ken Olsen did doubt the need for computers in the home, but the evidence is based primarily on the testimony of David Ahl who was perturbed when the personal computer project he championed at DEC was not supported by Olsen in 1974.
> Olsen’s resistance may have been similar to that expressed by another DEC executive, Gordon Bell. In 1980 Bell thought home terminals would act as gateways to remote computers which would provide appropriate services.
* https://quoteinvestigator.com/2017/09/14/home-computer/
It was supposedly said in 1977: most computers at that time were not small, and so it would not be surprising that people would not expect the general public to desire a large, power-hungry, noise-y apparatus in their house.
That's exactly the point. Until recently, AI models that could run on home machines were so bad that it was very hard to imagine anyone wanting to.
And, like the overly large machines of 1977, models are getting faster, leaner, and better. It's happening a lot quicker, though.
This is why I'm bearish on Anthropic, OpenAI, and friends. I am not confident that we will continue to see the same pace of improvement in frontier model capabilities as we have seen over the past year or two - not using similar mathematics at least. But I think that getting results that are close enough to the same standard to be a realistic substitute but in a model small enough to run locally may well happen quite quickly. And if it does - where is the moat to defend these AI organisations with their astronomical budgets when they're already starting to price more realistically and that's already killing a lot of the hype they've enjoyed until very recently? They have an accidental moat because they bought up the global supply chain for storage but that surely isn't going to last once the data centres to hold that storage are becoming liabilities.
If model performance asymptotes and CPU/GPU and RAM keep growing, even slowly, then eventually we will have frontier models on desktop that are totally competitive with hosted. It’s only a matter of time.
You already can if you’re willing to spend many thousands of dollars on a beast of a machine. I’m talking about middle tier desktops and laptops here. Maybe eventually even phones.
The only way hosted stays strongly competitive in that world is if they can keep pushing the frontier or by playing the classic social media and SaaS games of network effect building and integrations.
Many people might still use hosted, of course, but what I really mean is that their multiples won’t be justified and they will have little to no moat. AI will become commoditized, like a sophisticated next generation form of an encyclopedia with search.
We kinda ended up with terminals connected to mainframes anyway. The terminal being the web browser, and the mainframe being SaS. So it wasn't that far off.
the network is the computer
It doesn't really need this much explanation.
People take these quotes out of context all the time. Said in a business context, there was no need, at that time, for someone to have a personal computer.
There's no business justification in 1977 for a personal computer department at a business. It's similar to the gates quote about RAM (I think it was 64KB?).
These statements aren't meant to be forever quotes. Their business plan quotes.
> It's similar to the gates quote about RAM (I think it was 64KB?)
640, and Bill Gates said he either never said that, or at least never remembered having said it. I think there is no evidence anywhere that he did.
https://www.computerworld.com/article/1563853/the-640k-quote...
That exact quote? No, never. He said something like: current computers at the time had 64kb of RAM, so the OS was designed with a limit of 640kb, and he believed this would give them 10 years of future proofing. As it happened, that limit was reached much faster, in about 6 years.
Or maybe he simply made a mistake. Big deal. This doesn't speak negatively of his other achievements.
He had a long career and presumably many successes, and is fallible like the rest of us. But a half-remembered zinger with no context makes for zippier posts I guess.
The early popularity of Minitel, the continued popularity of ssh/tmux, and the web browser itself indicates that bespoke client applications are not the only way. He wasn’t directionally wrong.
The simple explanation is that predicting the future is generally impossible. It doesn't matter if it's Olsen or anybody else.
I will not be spending thousands in hardware to run the worlds most mediocre llms at meh speeds. Sorry. I know for llm bros they think every output made by an LLM is magic, like every NFT guy thought every NFT collection was game changing, but there's nothing useful you can do with llms and 128gb of RAM (and there never will be) unless you have llm psychosis. Who cares.
or "640K ought to be enough for anybody."
https://quoteinvestigator.com/2011/09/08/640k-enough/
Nobody ever said that, at least not as an assertion or prediction. The actual instances of similar language are from multiple people describing their earlier thoughts before they learned it wasn’t true.
There’s no public proof this has ever been said, and if it was, if it was not taken out of context.
I have that many browser tabs.
You seriously think running LLM is the same thing as general computing?
It’s better, it’s useful even for those who don’t have a deep knowledge of computers. I’d expect more AI users than programmers, than ms-word users, than excel users.
That’s too strong of an assertion.
Local models aren’t deterministically equivalent in capabilities to foundation models. Home computers are turing complete; just like a mainframe. They are just slower. Often not slower enough to matter.
Most people are ok with slower. An AI that lets you edit a family picture, in say 30 seconds, locally is preferable to one that is instantaneous but requires you to submit that picture to examination/storage/training/sale in someone else's AI ecosystem. If i want to crop my ex out of family photos, i should not have to first give that photo to Microsoft. If want an LLM to write a book report for me, i dont want it also alerting my school. And if i write a memo for a client, and i want an LLM to check the spelling, i dont want that memo leaked either.
It’s completely technically possible to have cloud services where customer data is opaque to the provider. Some of Apple’s services are like this already, for example.
I think there’s a sweet spot currently with munging your data blindly on the server so that your client device battery still lasts all day.
Meanwhile Apple and others push on with making client side models more efficient so that eventually the server costs and complexities go away.
dont want to share my pics with "cloud services"
I'd like to think so but the existence of Google and Apple and Microsoft's cloud based photo tools with phone integration suggests that's false.
You could run a pretty good home server on $50 of gear and yet we never saw any real adoption of OwnCloud/NextCloud style products as an alternative to Google Drive/Photos or Apple Cloud.
Why should LLM/Transformers be any different? Especially when you need a proper expensive GPU to run them instead of a Raspberry Pi?
Apple's photo tools run on device, and they'll probably ship more on device foundation models at WWDC too.
On-device AI is going to be important, I think. It doesn't have to take the form of a chatbot UI to be useful.
After the latest round of cloud storage price increases my non technical wife has been asking if we can do local backups instead...
> Most people are ok with slower. An AI that lets you edit a family picture, in say 30 seconds, locally is preferable to one that is instantaneous but requires you to submit that picture to examination/storage/training/sale in someone else's AI ecosystem.
Maybe if you ask them that question, but if you show them two products, they'll definitely prefer the faster one. 30 seconds is a long time to watch a progress bar.
Fast and public, or slow and private. Not everyone wants, or is allowed to, share their data with the AI world. And do not doubt that every bit shared with an AI service will be used for training.
The question here is about markets though. Not everyone wants x but if the vast majority of people want y, x is going to be niche and expensive.
You don't think the commercials of Google's AI photo features aren't going to have an impact on Apple users of their phones can do a worse version of that feature and it takes longer?
Plus there's the other question. If this thing is slower ... what's the price? The desktop/mini-pc version of this is $3000, after all. At this performance level what is an acceptable price for the laptops?
People definitely aren't going to accept more expensive + slower ...
He’s just a braggart. When you see something like this in somebody’s personal bio on social media, it’s basically a banner that means “take everything I say in the context of me promoting myself.”
Qwen 3.6 is far ahead of Gemma for most (but not all) things. I've deployed it out across a number of M5 MacBooks and it's genuinely useful for many tasks. It won't replace an Opus or current gen Sonnet sized model but it's still amazingly good for its size and probably as good as or just a bit before Sonnet 4 era. Far more reliable for tool calling, coding, agentic tasks and faster than the Gemma models especially with MTP.
Qwen 3.6 is a toy compared to DeepSeek V4 Flash or Pro. These models can now run on Apple Silicon hardware with as little as 32GB RAM for the Flash (with 2-bit quant, which is still quite capable) using SSD offloading, with just-about-reasonable performance for interactive use, and far better performance on longer contexts than Qwen (due to the more efficient KV cache/attention mechanisms in DeepSeek).
Very significant improvements may be viable for unattended inference via large-scale batches, which can reuse sparse experts and thereby mask some of the latency involved - this is quite unique to DeepSeek, again due to its efficient KV cache.
Qwen 3.6 27B still curb stomps Deepseek V4 in coding
1. Deepseek V4 is still in preview (training is not finished)
2. Qwen is much more demanding and borderline unusable on consumer hardware because it's a dense model. The 27B parameters are active all time for each token. It's not a MoE architecture where a router activates only some of them.
3. Qwen doesn't like quantization at all.
Deepseek V4 Flash still has 13B active params though? That is about half as many as Qwen3.6-27B (and much more than Qwen3.6-35B-A3B). Given that RAM (even on a base M4 or 'regular' Intel/AMD system) is like an order of magnitude faster than an SSD, even Qwen 27B running from RAM will be much faster than any Deepseek V4 model with SSD offloading. And the MoE will be much faster still.
Qwen 27B is also small enough to completely fit in a high-end consumer or mid-end pro GPU, like an RTX 5090 or Radeon PRO R9700. I found results claiming 30 tokens per second generation for 27B(-Q4_K_XL) on an R9700. I doubt you get more than 5 tokens per second doing SSD MoE streaming.
Even for relatively short contexts, I honestly already find the ~30B class MoE models to be only borderline acceptable in terms of speed on my laptop (Ryzen 7 7840U, 64 GB LPDDR5-6400), though I use Gemma 4 26B-A4B more than Qwen3.6 35B-A3B.
> even Qwen 27B running from RAM will be much faster than any Deepseek V4 model with SSD offloading.
If you have reasonable amounts of RAM to cache the most likely experts, that's not true at all. Qwen 27B is marginally faster on a nearly empty context, then falls behind as context length increases due to the different attention mechanisms. Prefill for Qwen is much faster, but you're still comparing vastly different model sizes and capabilities. DeepSeek Flash is the best deal overall.
> completely fit in a high-end consumer or mid-end pro GPU
Or you could fit the dense portion of a much more capable model and still take advantage of that hardware.
> the most likely experts
Is that how MoEs work? I though that an important constraint for MoEs is that experts need to be uniformly used to make sure they can be used effectively. If there is a 'common subset' that, if anything, sounds like a symptom of undertraining (i.e. the same trick will not work as well for Deepseek V4.1).
Also, even if your MoE hitrate is 90%, you still spend half your time waiting for the SSD, giving similar total speed to a 27B model!
Finally, it looks like Deepseek V4 is pretty much only runnable with antirez's ds4, and SSD streaming only works with Metal; but I would like to try what you say with llama.cpp which uses mmap to also potentially do SSD streaming. (I can maybe try the large Qwen3.5 MoEs?)
> as context length increases
What kind of context length do you consider reasonable, though? From what I know, all models (even frontier ones) start degrading once you pass a few hundred thousand tokens. So realistically, limiting context size might even improve quality, especially if you use token-efficient harnesses.
> Or you could fit the dense portion of a much more capable model and still take advantage of that hardware.
Your point about consumer hardware was that it would be "borderline unusable" when running Qwen 3.6 27B. However, you need much less hardware to run a 27B than DSv4 Flash. In addition, you can do the same 'trick' with low-end GPUs and small MoEs: my desktop with 32 GB DDR4-3200 and an RTX 2070 8GB can run the ~30B class MoEs at 20-30 tokens per second and similar speeds to my laptop.
> Is that how MoEs work?
For any given workload/session? Empirically, yes, that's what has been found across different models. There's quite a bit of predictability that makes caching helpful.
> Also, even if your MoE hitrate is 90%, you still spend half your time waiting for the SSD, giving similar total speed to a 27B model!
There are ways of masking some of that latency, though it requires some architecture-specific cleverness which is less directly applicable to a generic engine like llama.cpp.
> Finally, it looks like Deepseek V4 is pretty much only runnable with antirez's ds4, and SSD streaming only works with Metal
The llama.cpp folks are working on adding support, and the ds4 project is working on CUDA support for streaming inference, targeting the DGX Spark.
> From what I know, all models (even frontier ones) start degrading once you pass a few hundred thousand tokens.
DeepSeek V4 seems to do quite well on recall tasks even with large context. That's one plausible benefit of its compressed attention mechanism, compared to earlier models. Some degradation will likely still be there, but it's not necessarily obvious.
As for why people are calling Qwen 27B "borderline unusable" that may have to do with it being a dense model which makes for an increased compute intensity and pushes users towards discrete GPU platforms, since those tend to have the most compute overall as far as consumer hardware is concerned. I might agree that Qwen 27B is quite ideally tailored towards these platforms, but that does come with some limitations.
I have to disagree with most claims. I run Qwen3.6-27b at 260k context and 40-60 tok/sec. It handles most coding problems as well as Sonnet 4.6 under OpenCode on our production tasks. (As an experiment, I run the same prompts for the same issues in parallel for Qwen 3.6 and Sonnet 4.6 and usually see little difference in performance). I see zero degradation from quantization in practice.
Settings: RTX 5090, 5-bit weights (Unsloth), FP8 KV cache.
Last time I tried running large MoEs on this PC, they had inferior quality at 2-3 bits compared to much smaller dense models at 5-6 bits, and were slower anyway.
A 260k context (close to the stock maximum for Qwen, though it's possible to extend it) will take ~16GB RAM for storing the KV cache, barring quantization tricks which severely degrade quality. That's a whole lot more than what DeepSeek requires for a similar context length, and makes it infeasible to batch multiple inferences together. This used to be the status quo for consumer inference, in fact it still is for models like Kimi and GLM (which can sometimes be smarter than even DeepSeek V4 Pro!) but we can also do better nowadays.
You can run the 35B A3B model which is an MoE. Runs great on a 5090.
I've got a Qwen 3.5 running on a 12GB 3060 and it's dumb as a stump but still smart enough to get some useful work done. Since it's my daily driver desktop I havent jumped to 3.6 since last time I did I quickly ran out of vram and locked the desktop environment.
But yeah, the Qwen line is pretty impressive on commodity hardware.
I must be using LLMs very differently than y'all, because I can't think of a single thing I would rely on an LLM that's "dumb as a stump" to do for me.
To me, LLMs are for asking research questions + exploring design spaces + pointing at codebases to investigate bugs. And those all benefit from the model being as "smart" (in terms of both fluid intelligence and burned-in knowledge) as possible.
I'm guessing there exist problems where "intelligence past a certain point" doesn't matter, so these medium-sized models can match the performance of the bigger models. But what problems might those be?
Things that are tedious but simple but I'm unfamiliar with.
"Go add a gh action to compile and deploy this thing and run its tests" is one I've found it's good at. Yes I know how to make a gh pipeline but it's always a hassle to remember what goes where.
Cranking out unit tests is okay. It's good at summarizing things so it's not half bad at writing jsdoc/xmldoc comments.
Qwen suffers quantization a lot, rendering it borderline unusable.
The HN crowd is, by and large, not the target audience for his self promotion. I guarantee there is one and this is more or less effective.
I also don't get why this twitter user is linked here, versus all the news articles about this new hardware that have been everywhere over the past number of days.
> you may run some models locally if only from a cost perspective
I have a hard time believing running a model on a laptop will be cheaper than running it in a datacenter. Why wouldn't economies of scale apply here as with every other computation?
The datacenter setting has huge economies of scale for low-latency, just-in-time inference using extremely large models, but that's not the only viable use of AI. Batched, unattended inference of possibly smaller and weaker models, while theoretically viable in a datacenter setting, is far from the best use of that hardware. This is where local AI is at its best.
A laptop is really a pretty bad form factor to run LLMs. Worst cooling, more expensive memory that you cannot replace, resell value depreciating fast. It’s fine for tinkering, small scale research, and demos but it’s definitely niche.
The vision NVIDIA is selling is pure marketing IMHO
This is assuming that you'll be priced the fraction of computing that you consumed. But you are actually paying for their infrastructure, for the R&D (and also the computation that went into training the model) etc. It is not clear that, for your own small computations, this kind of costs are needed, but you will still pay your share in the investment the provider made so that they could serve everyone's computation needs.
But, currently ... you're not. AI companies are operating at a loss, and are being subsidized by their investors.
Local may or may not be cheaper than remote now, depending on the details, but the factors you describe won't affect the math nearly as much as they will once that subsidization ends.
In that analogy bigtech AI is currently investing in cleaner air for all of us? We _could_ breath it through their hose, but might as well breath it outside.
It's cheaper for the AI provider to use your laptop instead of their datacenter.
What "every other computation"? I seem to have a lot processing power at my disposal here, between my cell phones, laptops, gaming PCs, various other hardware devices.
You're going to need to analyze the problem much more deeply because it sound like the standards you are implicitly applying would result in "economically, everything should be centrally hosted" but that is clearly not the result that obtains. Even a modern mid-grade cell phone is no slouch; you may not be running a current-gen frontier AI on it but you certainly can do a lot of other rather intense things locally that would have been laughable 10 years ago, like suprisingly high powered games.
Because economy of scale isn't really the right metric here. A machine you were you were going to buy anyway essentially has a TCO of $0.
AI models will pretty undeniably affect your electricity bill; yes you already own the computer, but it will cost more to run it if it's doing inference!
To a point, but we're talking a laptop, not a server farm. Even if you're going fullbore wide open 24/7 that's about $150/yr in electricity bills at average rates. Not quite nothing but in terms of AI costs that's pretty close to rounding to zero.
> not even considering business security.
I suspect personal privacy and need to run AI workflows to handle the litany of administration tasks of a household will be what result in regular need for local AI.
Apple is already out front with this on a personal, individual level, but they are not obviously headed toward multiuser/family-level ~biz admin with a persistent server running local LLM.
> this guy puts this everywhere, gives me probably the inverse of what he is marketing for.
Do you think he's in mensa too?
The security aspect is the main driver why I’m seeing so many businesses investing in local hardware. They know the models aren’t as good (caveat that they also can’t run Chinese models) and that’s ok. Places that really care about security and data governance already aren’t on the bleeding edge. They wait for the nice stable lts version, they lock down dev machines in frustrating ways and have lots of IT admin layers.
But they also want to taste the sweet fruit of AI so the only way to do this that a CISO will approve is on local air gapped hardware. It’s a niche but still a billion dollar niche.
Microsoft is working on this with their new execution containers (https://github.com/microsoft/mxc)
> However, it will make decent machines to play video games."
Where you will need games to be rewritten for ARM to get full performance, just like on Apple's M series chips.
Maybe they just mean from a "it can run a lot of DLSS" perspective.
> "Ranked in the top 2% of scientists globally (Stanford/Elsevier 2025) and among GitHub's top 1000 developers"
This made me laugh. I can only image how insufferable this person is to deal with.
128GB seems the sweet spot for local models. I can program and install most GitHub projects with opencode and QWEN 32b with mtp.
anyone whose addicted to token theoughput is losing the operational knowledge and offline capabilities.
if you arent moving to the AMD 395 or MACs then youre hitching aride on the expensive calory ride
If you could buy a 256GB you’d be claiming that 256GB is a sweet spot. But I agree with you. Crack-tokens are not the future.
no, the fact that MACs and x86 and soon ARM are all going to have 128GB models in every sector, yeah, sure.
But watching everyone flounder because claude goes down or forcing you on API costs.
I'm programming things that'd take me days with a PC that, without OpenAI's VRAM shenagans, would cost you $2k.
It's more than just 'this is what I could do' it's definitely about 'this is what anyone could do with a new PC purchase'.
You must be unaware that System76 was already selling 192GB machines, mac studios used to be 512GB max. The only reason why we don’t have them anymore is that we are in RAM shortage.
Those 192GB aren't unified memory though. 128GB on Mac or 395 can be used by both CPU and GPU. It's the GPU + large memory that opens up fast local LLM inteference.
Yes, true. But if we had the ability to buy that much RAM in the laptop, everyone would be looking in that direction. Until this thing discussed here comes to the market, “we didn’t have computers with unified 128GB RAM either” (except of macs).
I'm aware you can have more. the term "SWEET SPOT" references a area that anyone/everyone can get to and isn't some magical expensive unicorn.
You're doing what the IT industry has been addicted to for decades: number goes up.
> You're doing what the IT industry has been addicted to for decades: number goes up.
No, I have a hands on experience with bigger models, and understand the advantages of using them.
you mean you're addicted to not understanding anything you do. That's fine. The rest of us arn't going to experience the glory of api bills going up.
You also probably believe you need to 'escape the permanent underclass'
You assume I use the subscription. There are other options but they require more than 128GB unified RAM. You also assume a lot about how I work. And those final assumptions about what and how I think of others speak more about your anxieties rather than what I think.
You assume a lot. Sometimes it’s good to simply ask a question.
Lots of people are already running AI locally. They are the people buying up all the consumer-grade nvidea gpus. What are they doing with them? Well, the same things people with home media or email servers are doing: stuff they dont want to share with the general public.
I want to reduce my dependency on companies like Google, OpenAI, and Anthropic. Aside from the concerns of data sharing I'm also not a fan of how they run their operations, for example Anthropic now using xAI's Colossus data center which is poisoning a marginalized community, or OpenAI getting in bed with the military.
Not everything I want to use an LLM for requires "PhD level intelligence", and increasingly I'm finding more uses that involve sharing my personal data.
Yesterday my local model helped me when looking for a doctor who is in-network for my insurance. I threw it a screenshot from the providers search results and it looked up reviews for all of them.
Which model are you running?
Qwen 3.6 35B-A3B and 27B both at Q8 on a Strix Halo machine
My local AI is currently upscaling an old british comedy from sub-dvd quality to 1k. (It is not availible other than on DVD.) It looks like it will take about a week for my pair of 5060s to chew through the task.
Which show?
Chelmsford 123
I own the DVDs so I'm OK upscaling/editing my own copies for my own use. But if I ran the task on an ai service I would no doubt trigger copyright issues.
DeepSeek Flash v4 is the leading local AI on 128GB machines, and DS4 is still in preview (training not finished), no?
Especially on Dwarfstar.
I hope a family-level AI appliance is a thing later. Local non-cloud assistant that lives in the house, families interact via voice or phones or whatever. Knows the contextual family stuff you need, etc.
We didn't get people buying family-level file servers for the family photo gallery and documents at any real scale, so i doubt we'll see similar for AI especially when the cost is that much higher for GPUs vs an SBC machine.
because nas hardware and software suck and everything else was a poorly executed subscription product...i think one was called helm, another was by early twitter alumni. imagine a home device that manages and maintains itself and is a joy to interact with.
And why would the hypothetical "OwnAI" product be any different?
not automatically, but a meaningful step up in ease of use (managing photo/video backup from all family devices) without a subscription would be a solid foundation
> "Ranked in the top 2% of scientists globally (Stanford/Elsevier 2025) and among GitHub's top 1000 developers" - side note but this guy puts this everywhere, gives me probably the inverse of what he is marketing for.
Lol yeah seriously, that stinks "I ask AI to generate a huge amount of bullshit and upload it to pad irrelevant stats".
Absolute loser.
I agree that it sends the wrong symbol, but actually Daniel is great. He cares tremendously about doing work that is actually real-world useful. I've co-written a few papers with him, and he's really hard working and open to outside suggestions. The danger is that if you send him comments, he'll eventually manage to rope you into writing a new and improved version. Seriously, if you are a non-academic computer scientist with a good idea that you want to publish, he'd be incredibly open to working with you.
As to why he now has this on his blog? I also cringe when I read it. I presume someone told him he should self-promote more, and this is his lame attempt to do so. He's almost certainly the most cited person in his department, but it's entirely possible that none of his colleagues actually know this. Cut him some slack. Self-promotion is not his strength. He's a nerd's nerd, and not a marketer. I'll mention to him that his attempt here might be backfiring when I'm next in contact with him.
I kind of get it in the sense that every academic has to make themselves somewhat comfortable with self-promotion even if they don't like it. It's an important part of getting funding, but putting a blurb like that everywhere just hurts his credibility I think.
I cringe calling it out but it just stood out as it was plastered everywhere and I actually have never seen his links before.
> As to why he now has this on his blog?
He doesn't just have it on his blog, he has it EVERYWHERE. Sometimes 2 or 3 times on the same page.
He's not a loser; he's done some really fun work that many people use daily. I've used his range mapping trick in multiple projects/papers. It's elegant.
It sounds like he's gotten bad advise about how to market himself /or/ this is being marketed to people who have bigger checks to write and whom he believes will be responsive to this kind of marketing. As an academic, it rubs me very wrong - I think it's detrimental to the field when we get into h-index stacking contests or citation count comparisons. But I don't know what incentives he's responding to, which seems important for putting this stuff in context.
(as an aside, it turns out that polars + fastexcel is about 10x faster than pandas + openpyxl for searching that dataset, if anyone else is curious what he was actually talking about. :)
I found his website, https://www.lemire.me/en/ , and the "2%" brag is the very first sentence, geez.
Being the top x% is what OnlyFans girls brag about, professor...
And it's not exactly brain surgery, is it? https://www.youtube.com/watch?v=THNPmhBl-8I
> Daniel Lemire’s blog is one of the top 50 most popular blogs on Hacker News, the standard tech news aggregation site.
Citation needed
https://refactoringenglish.com/tools/hn-popularity/
For posterity: It's rank 34 at the time of this comment
That lines looks very cringe indeed, but the guy has some crazy good blogposts on SIMD stuff.
I think the local-model use case is going to become less niche pretty quickly if the models keep getting smaller and more capable. Even if most people do not care about privacy or offline use, the cost argument is pretty strong
This feels fluff to me on the part of the author (whose work I don’t want to trivialize) but I don’t think they’ve actually looked deeper than a paper spec sheet on this.
1. Yes it has the same number of cores as a 5070 mobile. It’s also running at a shared peak of 2/3 the bandwidth and a shared peak of 2/3 the TDP. The GPU by itself will likely perform at half the dedicated units performance
2. Apple may not have SVE2 but they do have the AMX (private) and SME. I don’t see why he thinks the SVE2 will give him more performance than the SME.
3. He mentions a single core type but doesn’t mention the total makeup. We already have known for a year how the DGX Spark compares to Apple chips. For CPU it’s roughly equivalent to an M3 Pro and for GPU compute (not rasterization) it’s between an M4 Pro and M4 Max without considering bandwidth.
The real advantage to these is that they run CUDA. That’s it. Otherwise when they launch they’ll be 2-3 generations behind where Apple is and 1 gen behind AMD.
The other super power of the DGX Spark was the NIC for pairing them together. But that’s been removed here too.
Prefill is another advantage vs. Apple. It's way way way way faster on a spark than it is even on an m5 max.
Same model, same quant, same query, as close to as matched settings as I can get from vllm, and for workloads with large prompts + low cacheability, one of my sparks will often be done responding before the mbp is done with prefill.
> GPU compute (not rasterization) it’s between an M4 Pro and M4 Max without considering bandwidth
You are likely thinking about token generation which is dependent on memory bandwidth where Apple has an edge. Spark's GPU compute is way higher than even M5 Max (17 FP32 TFlops), around 2x FP32 TFlops... It's literally 6144 CUDA cores like desktop 5070, slowed down by slow memory and lower TDP (29.7 vs 31 FP32 TFlops on 5070).
That’s only if you consider FP32 specifically. On average the M5 Max will pull ahead for tasks like GPU raytracing (it’s currently the fastest mobile GPU for Blender rendering) and token generation and other things that benefit from the higher memory bandwidth.
I’d also mention that you’re comparing peaks which the RTX Spark won’t be hitting. The top TDP is less than that of the DGX Spark.
I just think anyone calling this a beast and a game changer are conflating/extrapolating from different form factors and constraints
Lemire is very narrowly interested in CPU SIMD so within that niche it may be interesting. As you said, overall the Spark is good but not great.
It is absolutely fluff, and the only reason this worthless tweet is on the front page of HN is that this audience has a habit of canonizing certain people, and treating each of their bowel movements as prophetic.
Guy suddenly became aware of a chip that the rest of the industry long knew about, seems completely unaware of the competitors, and posts about how it's a BEAST and will be a GAME CHANGER.
Like the DGX Spark was a game changer? Eh, it has mostly been a massive disappointment. An overpriced nvidia laptop isn't going to change the equation an iota.
The Qualcomm Snapdragon X2 Elite Extreme trounces Nvidia's chip in single core CPU performance. It beats Intel and AMD's best, too. It has unified memory. It's the only CPU in the same league as Apple's M-series in both CPU performance and power efficiency. And it's available in laptops today, not later this year. People are sleeping on Qualcomm.
Garbage operating system support. If you can’t do Linux support it’s a bit pointless because there’s two platforms for this that matter: Linux and Darwin.
Qualcomm is like AMD was for GPUs for like decades. Lots of announcements and people on the Internet are huge fans based on web pages they’ve read but if you try to make it work it’s a nightmare.
Snapdragon X Elite doesn’t work on Linux so it’s a pointless platform. Enthusiasts have made M1 work better. Literally have old Macs running rather than use Qualcomm.
It trounces ARM's old CPU design. The X925 used in this Nvidia chip is 2 years old. X930 or C1 has shipped with Mediatek Dimensity 9500 which is what the Snapdragon 8 Elite Gen 5 / X2 Elite should be compared to. Although Qualcomm still has a lead in performance, but it is increasingly shrinking.
But perhaps more importantly. Nvidia seems to be doing a lot better with its ecosystem. Nvidia has much better distribution channels and partners building on top of their PC Gaming GPU. It also have gaming developers relations that is unmatched by any in the industry.
Qualcomm has so far failed to execute this, both in PC and on there Server CPU side.
Microsoft is sleeping on Qualcomm with their lousy port of Windows to Arm processors…
> with their lousy port of Windows to Arm processors…
What's lousy about it? I use it daily and have zero problems.
I'm not sure they are sleeping. I have an older version and it can run games and other things just fine, its just over priced and not properly cooled. The driver/firmware support from Lenovo / Qualcomm is purely garbage. You're lucky to get a driver update to fix anything. For months it just overheated and video would start corrupting but that got fixed finally. You cant just go to Qualcomm's website and download new drivers even though it looks like you can - they really dont get how modern GPU's work on Windows - a driver updates to optimize for games is really something important because of how Windows is but the experience is pulling teeth. If the systems were Neo priced (500-700 USD) and had a cooling fan I'd be all on board with these systems. Right now, AMD with unified memory is just the better deal for the $1200 (2025) systems to run Windows and an average workload.
and is Qualcomm is sleeping on Linux?
Seems like not? Judging based on https://github.com/qualcomm-linux something is happening, although I can't say how much. They definitively seem awake at least.
The problem with these chips on Linux is that something has been happening for months but you still end up needing to download special editions of ARM Linux images to get these devices to work properly.
Some distros still need extracting Qualcomm firmware from Windows to get Linux to work properly. Audio remains a challenge, like x86 Linux decades ago. Apparently camera stuff works these days but produces images of subpar quality.
These issues also occur on normal Linux. My experience with my Lenovo+Intel laptop was that it took three months after release for the firmware to work properly (and the Nvidia drivers took much longer, but that's my fault for buying something containing Nvidia hardware). Intel managed to do what Qualcomm did in months rather than years.
I hope Qualcomm finally sorts this shit out, I really do, but with the prices of computers these days, I'm going to need to see quite the discount before I'll consider buying anything with a Snapdragon.
> but you still end up needing to download special editions of ARM Linux images to get these devices to work properly
This is a problem with Linux on ARM generally (Android has had it since inception), it's not a Qualcomm problem.
They run a hypervisor under the OS, and dont support actually running directly on the hardware, its very odd.
one of the biggest issue i see is the devicetree nonsense. It makes every single laptop and bios version very unique and requires a lot of housekeeping. There are also big chunks of work (as i understand it) to be done around hibernate and decent suspend support.
My experience (wanted to use x13s as daily sriver) is that there was good progress for about a year, until jhovold was leading the charge, but something expired and qualcom as far as i can tell forgot that some progress should happen on x1 and x8c as well as x2.
It feels deeply unfortunate that even with Windows on AArch64 requiring ACPI that it still doesn’t suffice for Linux, unlike on x86.
And I know a lot of that lies on the vendors, but it does feel unfortunate (from a standardisation/conformance/certification point of view) that Windows requiring it doesn’t make it easy to boot other OSes!
Yes, Ubuntu on the previous gen Snapdragon X is still trash.
10000000x this. They have been sleeping on Arm since windows phone. I just don’t see them ever having an original thought again.
They could have had a 128core arm chip by now.
They have original thoughts! It's just that those employees get squashed by other divisions or having to meet short term quarterly profits it seems.
There's also the whole giant trillion dollar company doesn't want to invest and let small ideas grow. They only focus on things that move the needle, which isn't much at the size.
Had Microsoft executed and invested, they could have made a come back imo in both search, mobile & hardware. Unfortunately major lack of leadership or they just don't want those areas.
No doubt the individuals have thoughts. It’s just corporate never actually does ANY of them. It’s a MBA mill.
Unless the chip was called Copilot, they are not thinking anything about it. If was called Copilot, they'd have already figured out how to shove it down your throat.
Qualcomm is a “fool me once, shame on you, fool me twice, you don’t fool me twice” kind of situation. So many horrible experiences in the past that people are going to be hesitant.
Qualcomm are trying harder now it seems. But it will take time to repair their reputation in the PC market.
They burned me with the first gen Snapdragon X Elite. Before the various laptops with it were out they promised Linux support. Here we are, years alter, still no fully OOTB support. Ironically, the GPU firmware were just mainlined in the kernel 4 months ago, but they still haven't done the same for the 1st gen X elite.
Tuxedo computers tried and didn't succeed either.
I will never buy Qualcomm again. I avoid them on phones as well by just buying Apple. They do not support their hardware beyond the release.
> I avoid them on phones as well by just buying Apple
To each their own, but I don't recall Apple ever mainlining any of their drivers on Linux. You're rightfully angry on the laptop side of things, but Apple is much worse than Qualcomm when it comes to open source support for their phones.
Qualcomm probably shouldn't have promised Linux support in the first place. Everyone seems to love Apple's hardware even though you're practically stuck with macOS. Had Qualcomm just stuck to Windows-only, they would've probably received a much better reception by the tech press.
Apple doesn’t sell general purpose computers outside of their own hardware so this doesn’t make any sense.
At least Apple tells you they don't support anything except their own OS, Qualcomm just pretends to offer support.
Can you say more? I don't have any memory of Qualcomm-related scandals(?), but I just read the news; I've never really been a user of their chips.
> Qualcomm are trying harder now it seems.
Not really, the 1st. iteration got stuck in legal land and other delays.
They hired a good number of smart people who know how to do open source. So they’re trying. We shall see if it works.
Is it well supported under Linux?
Qualcomm has been upstreaming Linux support for some of their chips but they're not working fast enough and I don't think the latest chips are there yet unfortunately.
I've been keeping an eye on the state of Linux on the first gen of X Elite and it's sad that the potential is not fully materialized outside WoA. Take a look at what peeps are going through:
https://discourse.ubuntu.com/t/ubuntu-concept-snapdragon-x-e...
No, not at all, those machines are currently unusable on Linux.
Too bad Qualcomm provides shit drivers for Linux, never updates any of their drivers (had a Samsung/Qualcomm phone with drivers years behind the equivalent Google Pixel phone), etc... They are the absolute worst actor in the entire computing world, don't care how fast their chip is.
> X2 Elite Extreme
I'll wait for the 365 AI Ultimate Professional Enterprise Edition: Origins version
> People are sleeping on Qualcomm.
Technically speaking, Qualcomm acquired Nuvia, which is where this came from and that company came from ex-Apple engineers wanting to do what Apple said no for their chips.
So it's almost same CPU design (origins).
> And it's available in laptops today
Is there a desktop version ? For real work ?
People aren't sleeping on Qualcomm, they're tired of Microsoft Windows as a janky ass OS.
Why do people care so much about single core performance? We are all professionals here and I bet most of our workloads are multi core. I get that these new arm chips from Apple and Qualcomm are great at one thing at a time, but for professional workloads high end x64 chips still cannot be beaten on the desktop.
What x86 chips have the same or higher number of cores in the form factors that these chips are available in and are also more performant?
Strix Halo is 16 cores. Intel Core Ultra 9 285HX is 24. Apple is 18. Qualcomm is something similar too but I can’t recall. NVIDIA is 20.
Until you get to threadripper/epyc or Xeon territories (completely different form factors and TDPs) the arm chips are ahead on both power and perf than the x86. And even when you get to those areas, arm is equivalent or out performs them as can be seen by the recent neoverse x3 and Vera benchmarks.
Nvidia Spark has 10 mid cores and 10 fake cores. Probably all the others are faster.
Single core performance is the biggest factor for most day-to-day use of a computer, the stuff I do on a laptop. It's more important than peak multi core performance for web browsing and games. I only care about multi core performance when I'm compiling, and I usually do heavy compiles on a remote machine rather than on my laptop.
I agree with you, but also:
outside of anything else, amdahls law means that as the parallel performance grows, we become _more_ limited by the inherently serial code, and thus single core performance, not less.
Given that single core performance is "harder" (can't just throw more cores/sockets at the problem), it's also critically important.
> Why do people care so much about single core performance?
Because that't the only part this chip excels.
People are comparing apples with oranges since ages.
Here is the press release for the actual machine:
https://nvidianews.nvidia.com/news/nvidia-microsoft-windows-...
I have been somewhat surprised at the lack of commentators observing that this is Microsoft and above all NVIDIA launching a device that is fundamentally at odds with the metered cloud model of AI.
When you look at the other announcements and murmurings (better offline BYOK for Copilot, talk of an unmetered AI future) I think it’s clear that these two firms understand that cloud-only AI is not sustainable or inherently in their interests. But their willingness to undermine OpenAI with a product like this is notable.
Yeah, "unmetered intelligence" was probably the most used phrase at MS BUILD this past week. They are going hard into local AI
I don't think you can interpret this as anything other than a sanctioned rebuke, right? Everyone has a strong visceral sense for what that means.
Copilot just got proper "offline" BYOK support, didn't it? Presumably that was one of the things they were talking about. Though I imagine that has something to do with the fact that Zed has supported that properly for months.
Maybe. Or they are simply hedging their bets.
I think most people are not understanding what this kind of laptop will provide.
Before we get local AI, we'll be using hybrid AI.
Running big models locally is unrealistic ($$$$$) but, if you imagine an Agentic Workflow where some bits run on the cloud and other smaller tasks locally, it's an amazing deal. You don't need Opus/Code/DeepSeek/Kimi/etc to do basic stuff that models like Gemma4:12b/Qwen-27b can do locally with much less latency.
Having a laptop where I can use a remote big model and combine it with 5 local domain specific models, is something I would love to do today. Imagine using OpenCode and you've a small model deciding which tasks run locally, then decides if you've a good local model for XYZ task or if we use a cloud model.
My main concern is: Is this hardware powerfull enough to allow local quick models switch? Unlikely but I hope I'm wrong
Given the incredible progress of local models, on present trajectory I think we see comparable levels of performance to frontier models in two years on 128GB unified RAM and 6-bit quantisation. Note how the frontier models are now hitting superior benchmarks with only 200,000 tokens. I think we still have a long way to go with distillation.
> “Our goal is to deliver unmetered intelligence to every home and every desk with Windows,” said Satya Nadella, chairman and CEO of Microsoft. “RTX Spark marks a real breakthrough towards that vision.”
I expect computers with this chip will be about $4000. If Microsoft can deliver on local AI models that can orchestrate Windows and have solid real world intelligence, that will be an inexpensive business purchase compared to pay as you go tokens. I'm excited to see how this plays out.
And who in 2026 is still anal-fixated on a "Windows" PC?
It's just a personal computer. It normally runs multiple operating systems just fine.
Windows PC sounds like people talking about tech who are either payed by M$, or embed pictures into Word documents to send them.
Nobody has to kill the fun those OS agnostic machine allow, by artificially bind them to a shitty OS.
Enterprise, of course. They probably buy more PCs than the rest of the market combined.
Even for personal use, I'd imagine the amount of people dual booting Windows and something else are a very tiny minority.
Saying "Windows PC" is a pretty reasonable way to distinguish between "made by Apple" and "made by someone else" because the market of PCs that aren't made by Apple and don't come with Windows is really, really tiny.
To be honest, this seems like a strange hill to take such an aggressive stance upon.
> And who in 2026 is still anal-fixated on a "Windows" PC?
I'm assuming it's just clarifying this isn't about Macs.
The term "PC" is ambiguous, since it can either refer to all personal computers in its original meaning, or to the IBM PC lineage that is mainly contrasted with Macs. Remember the famous "I'm a Mac, I'm a PC" ads.
When you just say "PC", people today genuinely don't know which meaning you are referring to. And "IBM PC" is antiquated, and "IBM PC clone" is even worse. So "Windows PC" is a pretty decent name.
Do you have a better suggestion? Because "Non-Mac PC" doesn't exactly roll off the tongue. If you say "Windows PC", everyone knows what you mean.
And it's not an "anal fixation", there's no need to be gratuitously insulting.
And this isn't a "Windows PC" in the traditional sense. The reason people run Windows in the enterprise (and for some desktop home uses like games) is still hardware and software compat.
I run it for work because we make windows programs. We use drivers that don't exist on Win-for-ARM yet. So to most people a "Windows PC" is an x64 Windows PC still. The risk for MS if compat isn't good enough for Windows-Arm64 is that people might as well shift from windows entirely if they need new software and harware anyway.
Hopefully anyone who wants to run anything other than Windows on an Nvidia-produced device has learned their lessons at this point. Although, a cursed Nvidia Hackintosh would be extremely funny.
For normal people, there are three computer operating systems: Windows, Apple, and ChromeOS. Nvidia isn't going with ChromeOS and Apple hates their guts, so Windows is the only normal operating system they can market.
Their marketing makes clear that these devices aren't the piddly Chromebooks that ruined the desktop experience for so many people (expensive Chromebooks were nice, but rare in practice).
Qualcomm promised Linux support, failed to deliver, and now anybody burnt by their promise won't want to buy their hardware again. If they promise a Windows PC, people won't have reason to complain when Linux or FreeBSD or SerenityOS won't boot on there. Given Qualcomm's failures here, Nvidia is probably doing the right thing.
> Although, a cursed Nvidia Hackintosh would be extremely funny.
I did this for years. We ran Resolve color correction suites with external chassis to place multiple Nvidia GPUs in it at a fraction of the cost of the shitty TrashCanMac that was available. Lots of people continued to use the 2012 Cheese Grater MacPro with its older CPUs. The only way to get modern (at the time) compute in a Mac was to use a Hackintosh. Since it wasn't for personal use, not having things like AppStore, Messages, Music, etc wasn't a big deal, so building a Hackintosh was easier.
I built one for personal prosumer use around the time of the 1080s that allowed me more machine for the dollar than Apple offered. Once the M-series chips came out and they were capable of what the Hackintosh was doing for me put me off of building anything newer.
Windows is dying a death by a thousand small, user-unfriendly decisions. This is genuinely sad because the technology underlying Windows is actually very robust and flexible.
So, the partnership is maybe natural, but not prospective. Also, note how Linux is getting popular among gamers. Of course, it's way behind Windows, but the direction of the change is clear.
I'm convinced that Nvidia is not primarily targeting the consumer market and that the ultimate goal for its CPUs is the server space. The company invests effort where the money is, and consumer products account for only a fraction of its total revenue. Maintaining a presence in the consumer market seems more like a way to avoid a complete pivot than a strategic priority.
A big push specifically for Windows ARM from Nvidia seems like relevant information.
> It's just a personal computer
Your x86 machines were, but these are ARM SOCs. Many of them don't even support UEFI, let alone the upstream Linux kernel.
Getting rid of UEFI is bad?
However bad UEFI is, it's still better than the fragmented ARM boot wasteland.
Can you quote where I said that?
Sir! I'm not an LLM.
I follow Daniel Lemire and like his contributions, I also understand that the HN thread was created for discussion purposes, but I'd really appreciate having a reference to the spec or a source to the claims made, either here on HN or on the tweet itself.
I dislike the cycle of propagating news and assuming that someone else double-checked it.
I think you’re asking for https://nvidianews.nvidia.com/news/nvidia-microsoft-windows-...:
“News Summary:
- NVIDIA RTX Spark powers the world’s first Windows PCs purpose-built for personal agents, featuring 1 petaflop of AI performance, industry-leading power efficiency, full-stack NVIDIA AI and graphics technology, and up to 128GB of unified memory.
- NVIDIA and Microsoft collaborate to deliver a native Windows experience for personal agents, including new security primitives and NVIDIA OpenShell to run agents securely on primary devices.
- RTX Spark lets creators, AI developers and gamers render ultralarge 90GB+ 3D scenes, edit 12K 4:2:2 video, generate 4K AI videos, run 120B-parameter LLMs with up to 1 million tokens context using agents locally, and play AAA games at 1440p and over 100 frames per second.
- Adobe is rearchitecting Photoshop and Premiere from the ground up for RTX Spark to deliver 2x faster AI and graphics performance.
- RTX Spark-powered slim Windows laptops with all-day battery life and premium displays, as well as compact desktop PCs available this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, with models from Acer and GIGABYTE to follow.”
I do not see how it is a 'beast of a' anything. It has 300GB/s memory bandwidth, barely above AMD Strix halo (256GB/s) with the same 128GB RAM and less than half memory bandwidth of M5 Max 128GB (614GB/s). Emphasizing memory bandwidth because most people interested in it I suppose are AI enthusiasts. Also, Windows.
They have a lot of software groundwork ahead of them to make an ARM CPU viable for any kind of desktop use outside direct inference or training usage too.
AMD has the advantage that their x86 machines run everything, Apple maintains the whole MacOS stack, while Nvidia can barely scrape together one Ubuntu release per Jetson generation, it's beyond embarrassing. Maybe they ought to put those agents they keep droning about to some actual work on their OS support.
The interesting part to me isn't really the Cortex-X925 vs AVX-512 comparison, but Nvidia trying to make the GPU the center of a Windows PC rather than an add-in card
>but Nvidia trying to make the GPU the center of a Windows PC rather than an add-in card
over the last decade, many software (especially the popular and industry standard ones) shifted to GPU accelerated design. it's a push before NVIDIA even tried to capitalize on that.
https://xcancel.com/lemire/status/2062880075117113739
I really hope this will have proper GNU/Linux support, otherwise it will end up the same way Qualcomm ARM PCs did.
guaranteed it will be like Qualcomm arm, it's a partnership with Microsoft after all. we may see a community project to make Linux work on it but it will not have an official first class support and many things will likely not work properly.
Is it really unified memory? AMD Strix Halo is "unified" but you still have to allocate memory separately for cpu vs gpu. Apple Silicon is true unified memory.
My understanding is that this is the limitation from Windows not from AMD SoC. There are several internet resources to "enable unified memory support" on linux eg [1].
As a side note, qualcomm chip set on Android has been doing this for years (like Apple) so it's not super unique thing. It's more like there was no need before.
[1] https://www.jeffgeerling.com/blog/2025/increasing-vram-alloc...
Even then the "reserved" section is a carve out guaranteed chunk to allow stuff that might need contiguous physical memory (display scan out buffers and page tables, for example) and similar.
The GPU can still happily use all the rest of the memory for other use cases - which tend to be the bulk of allocations anyway. Though there might be performance implications - for example "moving" buffer ownership to the GPU would need to evict CPU caches, and often 4k pages and tlb lookups can be a pretty inefficient situation for GPU-style accesses.
That's been pretty standard for any SoC for decades. And "differences" to apple's SoC are more implementation details.
yes, but more due to OS limitations than hardware. You can use their GTT which is then _true_ UMA where GPU can grab whatever it wants from the memory pool.
This isn't the first time we have UMA on the PC, btw. When SGI did their PC workstations, their 320 and 540 PC workstations had what they called Cobalt graphics chipset and crossbar with their IVC architecture. They bypassed AGP at the time completely. It was quite unique to see strict UMA on a PC. Haven't seen it since until these new systems we're seeing now on PCs and Mac.
That's a software question, not a hardware question.
Some software assumes pre-defined set-aside pools of memory reserved for video purposes, but the chip does actually have access to the whole pool.
For local models, the useful part is not just having 128GB attached to the package. It is whether the GPU can practically use that memory without the usual VRAM-style constraints
Strix halo is unified memory. The memory allocation set in BIOS is overridden by the operating system if it has the capability.
Memory bandwidth is what matters, unified or otherwise. Discrete GPUs don't have unified memory either.
> you still have to allocate memory separately for cpu vs gpu
That's an API issue not a hardware issue. Regardless, I believe the major APIs permit seamlessly sharing pointers at this point? (I have no experience doing that though.)
>AMD Strix Halo is "unified" but you still have to allocate memory separately for cpu vs gpu.
IIRC that's due to maintain BIOS and Windows (+games & apps) backwards compatibility, but memory access speeds are the same.
It is unified in the sense that the OS can dynamically assign memory to CPU and GPU. Apple silicon is not a alien tech that other silicon vendors cannot implement.
The M1 Max from 2021 has better memory bandwidth. The M3 Max can be specced to 128GB.
Nothing new here, apart from being able to use CUDA on a less power hungry system.
nb: poster is Daniel Lemire (https://lemire.me), who is very skilled in getting performance out of compute hardware (e.g. via simd, cache usage etc)
As he likes to share often, "He ranks among the top 2% of scientists globally (Stanford/Elsevier 2025) and is one of GitHub's top 1000 most followed developers. "
based on citations and github stars? or what's the context there?
I was adding further citation based on his own claims. Not sure what context is missing.
Still, Microslop has repeatedly proven their ability to slow everything down to a crawl no matter how powerful the hardware. If you want it to be fast, don’t use Windows.
The same chip that's been available in the DGX Spark for like 8 months now...why are we pretending like its the next big thing.
A beast if a Windows PC to do what? Run Trams, Excel, Outlook, and a browser all at the same time? We could do that just fine in 2010…
Just give it a year or two and Windows will drag that sucker into the mud and run everything just a sluggish as ever.
The idea that any hardware performance increase will be eaten up by terrible software is an evergreen. A computer that could serve as the single server for a medium size enterprise 20 years ago, is no longer able to serve as a desktop for a receptionist. I'm not even sure we're talking diminishing returns anymore, we're probably past the point of maximum yield and into the negative returns at this point.
I don't think this is going to get any traction in the general consumer world, even less relevant than Apple Vision Pro.
(HN reaction to Vision Pro back in 2024 is almost hilarious if not ridiculous, looking at it today. I knew it would be a flop and I was so right.)
Sounds good, but how much does it cost? Is this going to be an affordable laptop or $6000.
I mean fast 128gb of ram is like $2k so it’ll be $3k after overhead for just the ram portion of device.
How much is this supposed to cost, fully populated with 128GB of RAM? How much would this laptop cost?
It's not that the NVidia chip has that much RAM built in, after all. It's that it can address that much. RAM is sold separately.
The dgx spark is the same chip and those are in the low 3s to 5 range for most of them depending on manu, storage config, etc. The dgx sparks also have connectx 7 cards in them to support the 200gbps networking for RoCE.
So I would expect the mini PCs to come in less than the sparks. Laptops I assume will be close in price with the addition of all the other laptop stuff.
No. It’s all integrated. Not something you can buy separately or upgrade later.
See [1]. There's not 128GB of on-chip memory. "Integrated" memory in this context means that the GPUs and CPUs all use the same memory. There are on-chip caches, of course.
[1] https://www.nvidia.com/en-us/products/rtx-spark/
Ehhh.
It is all in integrated into one monolith “superchip” package. The 128gb of RAM isn’t going to be purchased separately or be upgradable. At least according to all indicators. Which is what I was responding to.
$4,000-5,000
> up to 6,144 state-of-the-art CUDA cores
A RTX Pro 6000 has ~24K 5th generation tensor cores, I'm guessing this would then be 1/4 of the count but 6th generation? Wasn't clear from the images.
What is more important than core count is how the caching architecture is laid out. They could lay out those 6k cuda cores in a layout which provides much larger blocks of cache to smaller number of cores. That would increase the memory bandwidth which would be better for inference.
Sounds like the memory bandwidth is worse though;
> The memory is not as fast as dedicated GPU memory, but it is cheap enough while delivering enough bandwidth to run AI models locally.
Also "cheap while delivering enough" certainly sounds like someone is trying to temper expectations. It sounds like something sitting in-between GPU+VRAM inference and CPU+RAM one, not as a step above/besides GPU+VRAM.
Having slower memory may not actually lead to lower memory bandwidth. The cuda cores can be broken up into compute complexes which larger blocks of memory directly attached to the cores. These could be filled with read operations from the bulk system memory. You can start executing and then page the next batch of data in while compute is working. For LLMs you don't have much random memory access, you can sequence your accesses in blocks.
If these chips become popular I am sure you will see LLM architectures taking advantage of the parallelism.
It's effectively the same as the GB10 in the DGX Spark (Blackwell architecture, 6,144 CUDA cores, perf-wise comparable to an RTX 5070).
I've found it very useful for running big models, but it's not a screaming powerhouse in terms of raw compute.
They are early versions, wait 4 years.
Intel's basic architecture keeps accelerators away from main system memory, unlike, for example, IBM's POWER architecture where the CPU and GPU are equal 'users' of memory. It's not a great breakthrough to suggest something different. The problem is - it's different, and not compatible with a lot, or most, or all, existing hardware. Also, there are some security concerns, as @stego-tech noted.
It’s an opportunity for them to start doing away with the whole ATX thing where owners had freedom to mix and match at their own pleasure.
They'll ship a welded-shut box that requires an activation key to power on. Users will get to pick color sleeve it uses though.
Says running local llms isn’t relevant. Than says it is decent for games, which is just correct if you compare any gpu remotely similarly priced. I don’t understand what is the point he is making
They are useless if RAM prices are this high. $800 laptops with maximum 8GB are currently the norm, Windows 11 can't run on them decently. No matter how fast the SoC is with overpriced RAM they are slow. Systems that can make good use of them with 64-128GB are not affordable anymore thanks to Nvidia and co. This is a smokescreen. They'll probably sell them packaged as compute modules anyway.
> Windows 11 can't run on them decently
Windows 11 can run just fine on 8Gb of memory, what cant is Google Chrome.
It's just a DGX spark with faster memory and a windows boot?
Is this essentially an Apple M-Series chip in concept?
while unified memory may offer better performance than unsoldered DDR system memory, it still won't be as great as 1.8TB/s bandwidth on high end consumer GPUs right now.
nvidias master plan may be making it the new normal to have "only" 400GB/s bandwidth, thus gatekeeping local model usage further behind "more memory but not as fast as the cloud can do it"
I think it’s an interesting theory but a bit too conspiracy theory-ish.
Nvidia just wants to sell stuff to everyone.
And I think for professionals doing local AI work, products like Strix Halo and Apple Silicon are a competitive threat.
A big part of maintaining the leading software ecosystem is ensuring you have competitive hardware for all your users.
I also think the RTX Spark product is relatively low effort for Nvidia. Grab a Mediatek CPU and slap an Nvidia GPU on the die. Sure, that’s oversimplifying it, but still.
128GB of unified memory is a dream come true for local LLMs. VRAM has been the ultimate bottleneck for developers.
The competitor for this NVIDIA CPU will not be the now old AMD Strix Halo, but its successor (launched recently), which supports up to 192 GB of unified memory. Thus 128 GB is no longer SOTA.
While this NVIDIA system is inferior from the point of view of the memory capacity, its main advantage is that the top models will have a bigger GPU, i.e. with 6144 or 5120 FP32 execution units, compared to 2560 for the AMD GPU (compared to the NVIDIA CPU, the AMD CPU has a better multi-threaded performance for legacy programs, and a much better multi-threaded performance for the applications that use AVX-512).
However, these top models with big GPUs will also be much more expensive than the competing AMD system, while also being much more expensive than a laptop or mini-PC with an equivalent discrete NVIDIA GPU (which has the disadvantage of having direct access only to a much smaller, even if faster, memory).
I don’t think there is much improvement in compute for the new strix halo revision. The next one supposedly adds rdna4 cores or similar and more memory channels
I have a 128 GB LPDDR5X machine. It's a great workstation laptop (which is why I got it) but the memory bandwidth is just awful if you're wanting to use it for AI. An old Epyc CPU will fair better both in terms of being able to run full sized larger models as well as having higher memory bandwidth, and that's not a recommendation to go that route either as it's still not worth it.
It could help with exploding external LLM costs. Interesting to see how the adaption will be, which will mainly depend on the price.
This is what makes it interesting to me as well
Mediatek and Nvidia the horsemen of abandoning hardware after a year. The Jetson family still left a bad taste in my mouth.
Qualcomm is too. They mainlined the GPU firmware for the X Elite 2nd gen, but still have not done so for their 1st gen X Elite which they promised full Linux support for and failed to deliver, and have now moved on pretending they never said that.
How dare you question the golden goose egg-laying algorithm for trillions in stock valuation!
good to know, hope the price will be affordable, having a pc becoming a luxury :)
I’m not sure if you’re aware but there is a supply chain shortage for pretty much everything needed for a PC that isn’t expected to be solved this year or next year. There is no way that can be affordable
or moreso, the available supply has been eaten by rampant speculation, and hyperscalers have overpurchased vs the datacenters they can actually get built and power
Certainly not in the year of our lord, 2026. Maybe in a few years though.
Don't want to be too harsh, maybe I'm missing something, but the CPU is at least 2 years old, internally it has been a complete shitshow and that's a minor hiccup when compared to the firmware and software situation.
It's an interesting "newcomer" and the more the better but calling this a "beast" and a "game changer" is ridiculous to say the least.
Then there is the price..
> The game changer is the unified 128 GB memory. That is the path Apple took years ago. Instead of separate memory for the CPU and GPU, everything shares a single pool. It is increasingly popular.
> The memory is not as fast as dedicated GPU memory, but it is cheap enough while delivering enough bandwidth to run AI models locally.
So, the reason "dedicated GPU memory" is fast, isn't because it's "dedicated"; it's because the types of memory built into GPU cards — GDDR and HBM — are designed for throughput over latency.
Which is to say, GDDR and HBM memory could be shared with the CPU in UMA while still being "fast" (for GPU use-cases.) In fact, the PS4/5 and Xbox 360 / One X / Series consoles have UMA architectures that use GDDR memory as their main memory, with no regular DDR memory to be found.
What I don't understand: why don't we see UMA architectures where there's both regular DDR and GDDR/HBM memory mapped into the address space of the CPU+GPU? That seems like the best of both worlds: you'd have some memory that's "tuned" for random-access CPU usage (regular DDR), and some memory that's "tuned" for streaming GPU usage (GDDR/HBM), but either type of memory can still be put to the use it wasn't "tuned" for, just with slightly-worse performance.
I guess you'd need to do a bit of software work:
1. a bit of work in the OS kernel / malloc library to get CPU workloads to "prefer" allocating DDR memory over the GDDR/HBM memory until they've exhausted DDR memory (or maybe not, if you just tell the kernel the GDDR/HBM memory is something like a zswap thinpool);
2. and a bit of work in supported ML frameworks, to teach them about a hybrid strategy between UMA "allocate anywhere, it's all the same" and NUMA "keep assets in VRAM if possible; if you spill assets to RAM, then they must stream into VRAM on access" (i.e. "at allocation time, allocate as if the system were NUMA, VRAM first then spilling to RAM; but at execution time, use the UMA codepaths, no need to copy RAM into VRAM.")
...but once that's done, it's done.
Yeah when laptops are shipping 8Gb and Microsoft is suddenly interested in native apps, nope.
Tech companies have strangled their own market.
Laptops shipping with less RAM is exactly the reason to be interested in native apps again. Every app being a chrome/EdgeWebView process is the problem.
Is this somehow satire? This is just the dgx spark with keyboard and monitor in a convenient format. Since it has more stuff, I'm sure that the price mark up will increase too.
Up to $5000 because why not?
With that money you can build a real PC with rtx 5090!
Not with 128GB (less OS) available to the GPU you can't. The unified memory is the point with this machine (and the dgx spark).
They announced RTX spark days ago. Why is this post linking to a "leak" tweet on the frontpage now?
A hardware company that propose to buy more hardware from them.
Must be a new business model.
....
Step into my office
Why ?
Because you are fucking fired
This is the RTX Spark [1].
The obvious comparison here is the M5 Max where you can buy a Macbook Pro with 128GB of also unified memory. Obviously CUDA cores are specific to NVidia so it's hard to directly compare but I've seen claims that the M5 Max is roughly equivalent to ~4000 CUDA cores. This obviously depends on workload and whether the CPU supports the precision you want to use (eg FP4).
The M5 Max has memory bandwidth of 819GB/s. The RTX Spark I believe is ~600. So it might be slightly better than the current generation of Macs but likely worse than the expected M5 Ultras of the new Mac Studios (likely Q3 2026).
For comparison, a 5090 has >20k CUDA cores and 1800GB/s memory bandwidth with 32GB of VRAM. The RTX 6000 Pro (at ~$10k) has 96GB of VRAM, same bandwidth and ~24k CUDA cores.
We have to see what RTX Spark systems sell for but the DGX Spark is in the Mac Studio price range (~$4k).
I do think Apple has a real opportunity here but there offerings aren't quite there yet. The M5 Ultras might be a really attractive option for local LLMs. I expect them to be in high demand.
[1]: https://news.ycombinator.com/item?id=48352939
> I've seen claims that the M5 Max is roughly equivalent to ~4000 CUDA cores
Who claimed that? The M5 is still a raster focused GPU, dedicated matmul blocks be damned. For some workloads that napkin math might work out, but for many others it's a wild overshoot. Time-to-first-token still favors CUDA, and real-world training workloads aren't getting anywhere near Apple Silicon.
All of the memory bandwidth in the world is useless if you spend 15 minutes processing 64k tokens worth of context prefill. This is where CUDA shines.
Related:
A powerful new chapter for Windows PCs, accelerated by Nvidia RTX Spark
https://news.ycombinator.com/item?id=48352693
Nvidia RTX Spark
https://news.ycombinator.com/item?id=48352939
Does this person know that this is the same GB chip in the DGX Spark? It isn't some proposed thing, it's a chip loads of people have on their desk right now, and there are endless benchmarks of it.
Decent single core (a long ways from Apple level, but decent), but it makes up for it in cores to provide M5 level performance, CPU wise. Memory bandwidth it is kind of starved, at 1/6th many GPUs.
They got Microsoft to customize Windows for the RTX Spark, and will likely have to brutally throttle it when running as a laptop (it's literally a 140W TDP chip), and that's neat. It's going to be a very expensive laptop.
This is probably the better way to frame it: not "Nvidia is proposing a new CPU system" but "Nvidia is trying to move an existing GB/Spark-class platform into a Windows PC form factor"
I heard the memory bandwidth is not just slower than on a GPU, as expected, but is significantly slower than Apple’s unified memory.
CPU/GPU is decent (800 GB or so), memory is slowish (300GB or so). Some Apple M are slower, some are faster.
Where did you get those numbers from?
DGX Spark has a maximum of 273 GB/s bandwidth in ideal scenarios (hard to reach)
That puts it between an M5 (153) and M5 Pro (307)
The 900 GB/s is from the NVLink-C2C interconnect, if you were wondering about that. They quote "up to 900 GB/s of bidirectional bandwidth between GPU and CPU".
Mind you thats not to/from memory, which indeed only has 273 GB/s.
Ah I see. But the only C2C equivalent on the Apple side is the UltraFusion which is 2.5TB/s if I recall correctly.
Yes its not an "Apple M killer" at all. Also, the available official performance numbers are partially overstated (1 Petaflop is only possible for sparse FP4 models, "in theory").
Perhaps a sobering rule of thumb: if it was actually useful, you couldn't buy them because someone would scoop them all up to shove them in a DC and make money with it.
Plus John Carmack has reviewed it, he was not amazed.
Not gonna lie, I'm buying one of the 128GB ram ones for local inference if price is human.
Wait a minute!
Nvidia going from GPU to CPU now?
Nvidia has been making CPUs for 10-15 years.
Will it support Linux?
Are their enterprise orders slowing down? Why use precious maxed out fab capacity on consumer stuff when it could be an enterprise chip?
It uses LPDDR5X instead of VRAM and will still sell for a premium while pushing their presence even further in every side of the AI market. This was one area AMD was ahead in and now Nvidia is probably better off making this to compete on that front while still being better off than making a 5090.
That doesn't answer the question. If the high margin enterprise GPUs are saturating the fab capacity you wouldn't expect them to be pushing this. But IIRC those all have oodles of integrated HBM at this point so I wonder if fab capacity for that has become a bottleneck.
I believe it does - the reasons why are exactly differences like LPDDR5X vs HBM3e. Not every fab is capable of making any type of chip another fab makes. If you can make a product with different chips and still sell it for a premium why would you not just because the fabs for your DC product's chips are busy?
Looking at it more, I believe the story repeats with the TSMC processes used for the CPU vs chips like GB200 as well.
Even if none of the above were the case, the question still isn't "why not make the enterprise GPU" it's "why not make the higher margin per chip area product". If the NV1/GB10 take less die space and cost a lot it's not immediately apparent the enterprise GPU actually nets Nvidia more $ per die or not. That's why it's relevant these will be sold at a premium.
It already is an enterprise chip. This is about Microsoft not having the equivalent of an M3 Max or whatever laptop.
And maybe for NVIDIA and MS it is also about them quietly betting that local models are, in fact, going to be good enough for most tasks pretty soon.
This is an enterprise offering. It'd take a guess its to try and stop the bleeding over to macOS. This launch, plus WSL containers, their own de-bloat winget config, mxc, etc. all seems like they're saying "pls stop leaving for macOS, see, Windows can be a great dev machine too."
This chip was designed before the shortages. I think they'll order just enough units to say they released it but not enough to put a dent in Rubin.
I am not sure how many people will run AI models locally. It still seems like a niche application to me.
I'd say this relates directly to the cost of running AI models remotely.
And we won't know what the actual cost will be until AI vendors recover the huge pile of cash they've dumped into development (plus interest).
I think it's niche now because getting the hardware to run it is expensive and the quantized models don't work as well. If those improve then it would be a no brainer to pay one off for the hardware instead of a fortune for API calls.
I am not really convinced that four bit quantisation is that bad; almost certainly six will be enough. But Google are making claims for their QAT tech in Gemma that they are surely using or testing in Gemini that it preserves nearly source model quality while reducing footprint.
The hardware for 50 tokens per second with a four bit quantisation of Gemma 4 26B or the sparse Qwen 3.6 is not really that expensive: it’s a secondhand M1 Max.
Beyond that, I agree. I think moving planning tasks to local is a now thing, not that it really has much impact on token spend. I also think many small coding tasks are fully within the grasp of the above two models.
The main issue right now is that the software landscape is rather confusing, but I reckon uncomplicated Gemma 4 26B QAT support with MTP is a few weeks away.
AI vendors are attempting to offer the whole apple. And they are spending huge sums of money in the process.
But most businesses don't really care about most of the apple --- they only need their special bite out of it.
For example, doctors mainly care about medicine. Nvidia is attempting to provide the hardware needed for local, specialized models.
I think it is likely to appeal to video and photo editors who want to use AI tools (the press release has a quote from Blackmagic Design, as well as from Adobe, who I think have no stomach for their own cloud AI).
But I don’t know about specialised: this could run quite large models with MoE.
Performances of local models are pretty bad compared to what AI vendors offer, token generation is just too slow to be that useful. And you need to allocate GBs of memories, something that will stay very expensive to buy for a long time.
Running local models will stay niche for a while, unless we see breakthroughs
Dumb idea --- how about if we limit local models to specific domains --- medicine for example.
Most doctors don't care much about engineering or accounting or software development or 10000 other things that big vendor models address.
This area is yet to be really explored. Nvidia aims to provide the hardware to do so.
That's a fairly obvious idea, not dumb at all, but unfortunately it doesn't seem to pan out. Trying to specialize an LLM in one area harms its 'cognition' in all areas. For instance, if you train a coding model without all the Shakespeare and soap operas and Wikipedia and pirated Stephen King books and ancient Roman history and whatever, you end up with a worse coding model.
I'm not sure anyone really understands why.
cant wait til someone figures how to run Linux on one of these
And it will be expensive - right?
Nvidia is milking the market now. We need more competition again - currently we have a mafia control the prices, not just Nvidia but all the AI companies. The price increases should be paid for them, not by us. "Free market" is being manipulated by them here.
"Major banana producer suggests shifting more ice cream store menus to banana splits, and increasing the amount of bananas per serving"
> I am not sure how many people will run AI models locally. It still seems like a niche application to me.
Bill Gates had a quote some years ago...
People have still not learned how fast we improve our tech and how much cheaper thing gets I guess :)
Memory isn’t getting cheap soon, and you need a lot of it for local models
All depends. The current technology will be cheaper in a year or two. The best cutting edge stuff will properly be even more expensive. But in 10 years time... we can run current SOTA models (or models that are equally good ) on our local hardware
Ah yes, if you count in decades, for sure I expect to run them locally
We had a thing called globalism that drastically reduced costs. Globalism right now is on life support. Given geopolitics, I don’t see how it’s going to survive.
"I am not sure how many people will run AI models locally. It still seems like a niche application to me."
Clip me :). You are currently living through the final stages of unrestricted computing in the hands of the 'public'. Our regimes are going to pull up the drawbridge in the name of 'safety'. Download the open models asap and prepare for an airgapped computing environment. That will be your frontier in not extremely neutered AI in the near future.
I am so hoping I'm completely wrong on this btw.
They have bought the governments, so what you model will probably become true. They just inflate the prices right now.