Nvidia is proposing a beast of a CPU system for Windows PCs

https://twitter.com/lemire/status/2062880075117113739

by tosh • 2 months ago

332 557 2 months ago

twitter.com

cyberziko• 2 months ago

good to know, hope the price will be affordable, having a pc becoming a luxury :)

crims0n• 2 months ago

Certainly not in the year of our lord, 2026. Maybe in a few years though.

dgellow5• 2 months ago

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

8note• 2 months ago

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

YasuoTanaka• 2 months ago

128GB of unified memory is a dream come true for local LLMs. VRAM has been the ultimate bottleneck for developers.

avocadoking• 2 months ago

It could help with exploding external LLM costs. Interesting to see how the adaption will be, which will mainly depend on the price.

adrian_b2• 2 months ago

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).

christkv• 2 months ago

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

zamadatix3• 2 months ago

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.

SwtCyber5• 2 months ago

This is what makes it interesting to me as well

jqpabc1234• 2 months ago

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).

chpatrick• 2 months ago

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.

jqpabc1234• 2 months ago

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.

dofm6• 2 months ago

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.

dgellow5• 2 months ago

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

jqpabc1234• 2 months ago

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.

tosh2• 2 months ago

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)

infecto4• 2 months ago

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. "

tosh2• 2 months ago

based on citations and github stars? or what's the context there?

tempodox• 2 months ago

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.

2OEH8eoCRo0• 2 months ago

Are their enterprise orders slowing down? Why use precious maxed out fab capacity on consumer stuff when it could be an enterprise chip?

zamadatix3• 2 months ago

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.

fc417fc8025• 2 months ago

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.

dofm6• 2 months ago

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.

thewebguyd8• 2 months ago

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."

wmf6• 2 months ago

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.

easygenes3• 2 months ago

Mostly a strategy move to protect the CUDA moat… Apple would take over mobile inference in a clean sweep without competition.

llm_nerd2• 2 months ago

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.

Apreche• 2 months ago

I heard the memory bandwidth is not just slower than on a GPU, as expected, but is significantly slower than Apple’s unified memory.

MrBuddyCasino4• 2 months ago

CPU/GPU is decent (800 GB or so), memory is slowish (300GB or so). Some Apple M are slower, some are faster.

MrBuddyCasino4• 2 months ago

Plus John Carmack has reviewed it, he was not amazed.

SwtCyber5• 2 months ago

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"

seanalltogether• 2 months ago

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.

joe_mamba• 2 months ago

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.

ankurdhama• 2 months ago

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.

ApatheticCosmos• 2 months ago

Strix halo is unified memory. The memory allocation set in BIOS is overridden by the operating system if it has the capability.

eigenspace• 2 months ago

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.

fc417fc8025• 2 months ago

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.)

flakiness• 2 months ago

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...

kimixa• 2 months ago

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.

SwtCyber5• 2 months ago

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

Keyframe• 2 months ago

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.

glitchc• 2 months ago

Memory bandwidth is what matters, unified or otherwise. Discrete GPUs don't have unified memory either.

sisve3• 2 months ago

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 :)

chaostheory• 2 months ago

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.

sisve3• 2 months ago

You are right. I haven't considered that enough. I do agree with you that globalism will be decreased even in the context of a decade, but I think it will be on a level that still give us some cost reduction. But I do think (at least hope) geopolitics will be better in 5-7 years.

dgellow5• 2 months ago

Memory isn’t getting cheap soon, and you need a lot of it for local models

sisve3• 2 months ago

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

infecto4• 2 months ago

"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.

iLoveOncall2• 2 months ago

"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.

netsharc• 2 months ago

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

nkurz2• 2 months ago

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.

SkiFire13• 2 months ago

That lines looks very cringe indeed, but the guy has some crazy good blogposts on SIMD stuff.

dgacmu• 2 months ago

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. :)

SwtCyber5• 2 months ago

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

unmole• 2 months ago

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?

dgellow5• 2 months ago

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

wazdra• 2 months ago

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.

jerf• 2 months ago

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.

TylerE2• 2 months ago

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.

itishappy• 2 months ago

It's cheaper for the AI provider to use your laptop instead of their datacenter.

zozbot23415• 2 months ago

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.

lrae• 2 months ago

Does it apply for every other computation? Purely for the computation part? You can host all kinds of things locally cheaper right now than in the cloud, no? (At least pre memory price hikes.) It does, of course, come with its downsides like availability/reliability, less convenience, scaling options,..., but purely the computing price - I don't see why it wouldn't be cheaper in the future - at least for some use cases.

sandworm1015• 2 months ago

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.

Zetaphor3• 2 months ago

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.

root-parent• 2 months ago

"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"

AaronAPU• 2 months ago

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.

throw0101a2• 2 months ago

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.

joering2• 2 months ago

or "640K ought to be enough for anybody."

fg1374• 2 months ago

You seriously think running LLM is the same thing as general computing?

dakolli• 2 months ago

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.

voidfunc• 2 months ago

"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.

flatline• 2 months ago

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.

bespokedevelopr• 2 months ago

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.

thewebguyd8• 2 months ago

Microsoft is working on this with their new execution containers (https://github.com/microsoft/mxc)

cyanydeez4• 2 months ago

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

throw12345678916• 2 months ago

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.

smcleod• 2 months ago

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.

Pxtl5• 2 months ago

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.

zozbot23415• 2 months ago

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.

epolanski4• 2 months ago

Qwen suffers quantization a lot, rendering it borderline unusable.

unstatusthequo• 2 months ago

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.

Pxtl5• 2 months ago

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.

jayd165• 2 months ago

Maybe they just mean from a "it can run a lot of DLSS" perspective.

falsemyrmidon• 2 months ago

this guy puts this everywhere, gives me probably the inverse of what he is marketing for.

Do you think he's in mensa too?

jb1991• 2 months ago

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.”

GeekyBear3• 2 months ago

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.

epolanski4• 2 months ago

DeepSeek Flash v4 is the leading local AI on 128GB machines, and DS4 is still in preview (training not finished), no?

Especially on Dwarfstar.

strictnein2• 2 months ago

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.

latch• 2 months ago

I also dislike his self-promotion, but his work _is_ well know and, as far as I know, well looked upon. I think he has more expertise and knowledge in this area than most (including what you'd find in the news).

bredren• 2 months ago

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.

alberth• 2 months ago

Is this essentially an Apple M-Series chip in concept?

trynumber9• 2 months ago

No, it stems from a lineage of Tegra chips that pre-date the M-Series. This chip was called GB10. One of its predecessors, GV10 was shipped in 2018. It was a 256 bit, unified-memory system on a chip with a Volta GPU and 12 ARM Cores. GB10 is a 256 bit, unified-memory system on a chip with a Blackwell GPU and 10+10 ARM cores.

BoredPositron2• 2 months ago

Mediatek and Nvidia the horsemen of abandoning hardware after a year. The Jetson family still left a bad taste in my mouth.

burnt-resistor2• 2 months ago

How dare you question the golden goose egg-laying algorithm for trillions in stock valuation!

thewebguyd8• 2 months ago

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.

SwtCyber5• 2 months ago

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

QQ003• 2 months ago

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.

cryo32• 2 months ago

Yeah when laptops are shipping 8Gb and Microsoft is suddenly interested in native apps, nope.

Tech companies have strangled their own market.

thewebguyd8• 2 months ago

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.

AmazingTurtle• 2 months ago

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"

dangus• 2 months ago

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.

dofm6• 2 months ago

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.

tantalor• 2 months ago

Maybe. Or they are simply hedging their bets.

gscott• 2 months ago

They get to keep more Ai usage from big providers, send telemetry, and inject ads.

thewebguyd8• 2 months ago

Yeah, "unmetered intelligence" was probably the most used phrase at MS BUILD this past week. They are going hard into local AI

dofm6• 2 months ago

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.

Yokohiii• 2 months ago

For me this is a push to segment the market into consumer and industrial grade RAM. Even NVIDIA and MS are not stupid enough to think they can keep going with RAM prices exploding. Consumers need hardware to subscribe to their AI stuff.

LLMs will get bigger and even with 128GB (that many wont saturate), you wont run future frontier models. For LLM vendors and integrators it's a handy thing to move lower quality inference to the consumers.

Also running local doesn't have to mean that the models have open weights. MS will likely start to distribute closed models at scale once the hardware is there.

Waterluvian• 2 months ago

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.

burnt-resistor2• 2 months ago

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.

thrance• 2 months ago

Will it support Linux?

ChrisArchitect• 2 months ago

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

PedroBatista• 2 months ago

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..

dagmx7• 2 months ago

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.

llm_nerd2• 2 months ago

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.

trympet• 2 months ago

Yes. This reads like a LinkedIn post rehashing old news. I’m not even in the industry.

storus• 2 months ago

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).

dagmx7• 2 months ago

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

wmf6• 2 months ago

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.

oofbey• 2 months ago

I cannot fathom why he brings up CPU SIMD as a potential comparative weakness on the NVIDIA Spark when it has teraflops of CUDA sitting right there.

cthalupa7• 2 months ago

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.

GuestFAUniverse• 2 months ago

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.

bigyabai8• 2 months ago

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.

rvba• 2 months ago

Getting rid of UEFI is bad?

zdragnar• 2 months ago

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.

jayd165• 2 months ago

A big push specifically for Windows ARM from Nvidia seems like relevant information.

jeroenhd3• 2 months ago

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.

dylan6043• 2 months ago

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.

kmac_• 2 months ago

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.

alkonaut• 2 months ago

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.

crazygringo• 2 months ago

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.

Aperocky2• 2 months ago

Well, there's the other problem, windows sucks.

I prefer Windows XP, or even Windows Vista, to Windows 11 with its copilot. And it's been a downhill race, even macs are more of your own personal machine than Windows today, which is saying a lot.

PC should be a PC, Windows is as they advertised, a Copilot PC.

pjmlp6• 1 month ago

The 80% of desktop market, and Valve porting Windows PC games into SteamDeck.

sometimelurker• 2 months ago

cant wait til someone figures how to run Linux on one of these

easygenes3• 2 months ago

That happened a year ago when these shipped as the DGX Spark with only Linux pre installed.

embedding-shape3• 2 months ago

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.

gravypod2• 2 months ago

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.

embedding-shape3• 2 months ago

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.

npn• 2 months ago

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!

thewebguyd8• 2 months ago

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).

ozgrakkurt• 2 months ago

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

jmyeet3• 2 months ago

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

bigyabai8• 2 months ago

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.

sherazp995• 2 months ago

Wait a minute!

Nvidia going from GPU to CPU now?

wmf6• 2 months ago

Nvidia has been making CPUs for 10-15 years.

userbinator• 2 months ago

Much longer than that...

https://theretroweb.com/chipsets/182

https://www.nvidia.com/en-us/drivers/uli-m6117c/

TiredOfLife2• 2 months ago

But this one has mediatek cpu with off the shelf arm cores

modeless3• 2 months ago

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.

bradfa2• 2 months ago

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.

derefr3• 2 months ago

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.

thewebguyd8• 2 months ago

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.

re-thc2• 2 months ago

Qualcomm are trying harder now it seems.

Not really, the 1st. iteration got stuck in legal land and other delays.

Danox• 2 months ago

Microsoft is sleeping on Qualcomm with their lousy port of Windows to Arm processors…

reactordev3• 2 months ago

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.

criticalfault3• 2 months ago

and is Qualcomm is sleeping on Linux?

hypercube33• 2 months ago

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.

sedatk• 2 months ago

with their lousy port of Windows to Arm processors…

What's lousy about it? I use it daily and have zero problems.

darkwater• 2 months ago

Is it well supported under Linux?

modeless3• 2 months ago

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.

diabllicseagull• 2 months ago

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...

izacus2• 2 months ago

No, not at all, those machines are currently unusable on Linux.

dismalaf2• 2 months ago

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.

SecretDreams• 2 months ago

People aren't sleeping on Qualcomm, they're tired of Microsoft Windows as a janky ass OS.

ksec• 2 months ago

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.

re-thc2• 2 months ago

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).

alt2272• 2 months ago

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.

hulitu2• 2 months ago

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.

Remnant44• 2 months ago

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.

dagmx7• 2 months ago

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.

modeless3• 2 months ago

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.

pjmlp6• 1 month ago

It is a feel good kind of thing usually repeated by folks that don't understand OS scheduling algorithms.

Unless we're talking about RTOS, threads are always interrupted, thus how fast a single one can race is irrelevant in the whole picture.

hulitu2• 2 months ago

And it's available in laptops today

Is there a desktop version ? For real work ?

Razengan• 2 months ago

X2 Elite Extreme

I'll wait for the 365 AI Ultimate Professional Enterprise Edition: Origins version

arjie• 2 months ago

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.

KeplerBoy• 2 months ago

How does Darwin matter more than Windows?

TiredOfLife2• 2 months ago

It has Linux support from wsl2

satvikpendem2• 2 months ago

Alternatively, just use Windows.

someguyornotidk• 2 months ago

Yep, very much this. I don't bother looking to them for anything in this space because the hostility they show towards general Linux support appears to be rooted in principle and appears to be very much deliberate. It almost feels like Linux support would run counter to how they want their processors to be used.

Whether this is true or not, it's pretty safe to assume anything based on their stuff is not for me.

danslo• 2 months ago

Support is slowly landing... https://www.phoronix.com/news/Lenovo-Yoga-Slim-7x-Gen11-Linu...

emsign• 2 months ago

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.

alt2272• 2 months ago

Windows 11 can't run on them decently

Windows 11 can run just fine on 8Gb of memory, what cant is Google Chrome.

fg1374• 2 months ago

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.)

monster_truck• 2 months ago

Wild that it took me so much scrolling to find some sense!

Spark DGX also remains a nothingburger, I would be livid if I spent this kind of money and had to waste time chasing down power cap bugs or A/B/C testing each firmware version to find the one that is least slow and also does not fail https://dredyson.com/the-hidden-truth-about-dgx-spark-perfor...

PeterStuer• 2 months ago

"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.

shevy-java2• 2 months ago

They have bought the governments, so what you model will probably become true. They just inflate the prices right now.

shevy-java2• 2 months ago

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.

derefr3• 2 months ago

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.

Rohansi11• 2 months ago

Theoretically, maybe? But they are completely different interfaces so it would surely get complicated. It's also approaching the current behavior in non-unified memory systems where you have two pools of memory with different performance characteristics. You'll realistically want the CPU to always use low latency memory and the GPU to use high bandwidth memory with very little moving between them.

amacbride• 2 months ago

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.

adamnemecek• 2 months ago

They are early versions, wait 4 years.

stego-tech7• 2 months ago

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).

Asmod4n• 2 months ago

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.

stego-tech7• 2 months ago

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.

jayd165• 2 months ago

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.

BoredPositron2• 2 months ago

gen3 is 16 years old.

vlovich123• 2 months ago

(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.

stego-tech7• 2 months ago

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.

b112• 2 months ago

But... it's rust.

jmyeet3• 2 months ago

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.

zozbot23415• 2 months ago

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.

woodson• 2 months ago

5090 ($2k MSRP but realistically $3-3.5k)

These days, more like >$4.1K (at least in the US).

dahart• 2 months ago

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?

Salgat2• 2 months ago

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.

simonebrunozzi• 2 months ago

What should Apple do, in your view, to "embrace" it?

Melatonic9• 2 months ago

It's also ECC ram but to be fair - yes quite overpriced. The RTX Pro line are basically what the Titan line used to be but way way more expensive.

Retr0id2• 2 months ago

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!)

bobmcnamara• 2 months ago

Oh no now I have to worry about shaders row hammering my OS ram /s

supertroop• 2 months ago

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.

p_l4• 2 months ago

SGI O2 was the famous "unified memory architecture" graphics system, two years before i740 that didn't really do UMA.

O2 was popular in systems where large textures or textures generated dynamically (like mapping external video input to texture) was important

seemaze2• 2 months ago

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.

canyp• 2 months ago

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...

aabdi2• 2 months ago

If this thing only has as much gpu bandwidth as the spark, it’s kinda pointles

cthalupa7• 2 months ago

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.

pbalau2• 2 months ago

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.

Gareth3212• 2 months ago

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.

ImprobableTruth• 2 months ago

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.

saltcured• 2 months ago

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.

Rohansi11• 2 months ago

Marketing, mostly? But perhaps also more flexibility with how much memory the GPU can directly access without reserving it.

cthalupa7• 2 months ago

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.

Salgat2• 2 months ago

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.

zozbot23415• 2 months ago

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.

testing22321• 2 months ago

The Unified Memory pool is the “game changer”

M1 knocking from 2020.

Gamed changed, past tense, six years ago. This is catch-up.

bombcar• 2 months ago

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.

jandrese• 2 months ago

Hell, SGI O2s from 1996 had this. For all of the hype the performance gains were pretty modest.

Rohansi11• 2 months ago

FYI it existed long before that. Shared memory between CPU and iGPU has been a thing for a long time.

zokier2• 2 months ago

AMD Fusion knocking from 2010.

nalekberov• 2 months ago

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

merb• 2 months ago

LPCAMM2

cm21872• 2 months ago

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)

MBCook5• 2 months ago

Is that required or just a choice Apple made?

Onavo• 2 months ago

Are there no PCIe standards that are sufficient to support both use cases?

What happened to PCIe 8 and CXL?

arka2147483647• 2 months ago

There is LPCAMM2, if manufacturers want to use it.

So, it does not have to be soldered.

to11mtm3• 2 months ago

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.

bpavuk• 2 months ago

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

GeekyBear3• 2 months ago

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.

ValentineC• 2 months ago

I wish manufacturers could consider a hybrid approach. There should be no reason an architecture can't support both unified memory (effectively L4(?) cache), and cheaper, upgradeable system memory on sticks for old-school application use.

nine_k2• 2 months ago

Maybe I won't care about upgradeability right now. The architecture is clearly in flux, the roles of traditional "CPU" and "GPU" are rapidly evolving. Maybe in 5 years, or even 3 years, a brand-new machine from 2026 won't be worth upgrading for a new role due to a seriously different architecture, but would only be relegated to do something "traditional".

pjmlp6• 1 month ago

A note that this was rather common on the days before PC clones took off.

The vertical integration many associate with Apple, was the common approach to most 8 and 16 bit home computers.

Naturally after all these years, many PC vendors want their margins back, and thus the phenomenon of everyone going back to vertical integration, especially in form factors that are ideal for such, like laptops, tablets and phones.

So the option boils down to classical desktops, or being picky on which laptops to buy.

tjoff• 2 months ago

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.

Joel_Mckay• 2 months ago

Jensen Huang has publicly stated he wants a future where "AI" agents use more PC computers than people.

It is ambitious, and absurd... like all CEOs that eventually go loopy. =3

david-gpu• 2 months ago

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.

Izikiel43• 2 months ago

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.

NikolaNovak• 2 months ago

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 :-(

maccard5• 2 months ago

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.

BuyMyBitcoins• 2 months ago

How much of that difficulty comes from the chosen game engine? I assume the engine is the primary factor in how resources are allocated.

Stevvo2• 2 months ago

What? It's incredibly easy to take full use of memory bandwidth. For example, put proper volumetric smoke/fire/explosion sim in your game. But game developers don't do that because they are lazy.

stego-tech7• 2 months ago

I live with a game dev myself, so I get it. Hell, it's hard even for PC developers who want to do things without leaning on abstraction layers or existing engines. Managing multiple discrete memory pools, asset swaps or calls between them, getting the respective subsystems to exchange data at just the right time so as not to impact other code and drag down performance - it's fucking hard in general.

A unified pool of memory suddenly makes that simultaneously easier, but also far more flexible, which frees up developer time and bandwidth to focus on other, more important tasks.

GTP• 2 months ago

While I'm a supporter of Rust, I have to point out that Rust's memory safety doesn't help against side-channel attacks.

up2isomorphism• 2 months ago

This kind of post shows you have little idea why cpu and gpu are not sharing memory in the first place.

jorvi• 2 months ago

Yeah, no. GDDR is functionally very different than SDRAM.

GDDR tries to push out as much bandwidth as possible, because that really matters for (traditional) GPU workloads. A constant but insignificant (= correctable) error rate is considered completely fine for GDDR, because that sacrifice allows the memory to be pushed much farther.

Meanwhile most (traditional) SDRAM workloads don't give a hoot about bandwidth but really care about latency. And ideally you want no errors, hence ECC RAM being so venerated.

If you unify memory, you're gonna have to choose to sacrifice one of those workloads or go suboptimal for both.

Weirdly enough this mostly matters for non-gaming workloads. The Apple M-series are absolute monsters in gaming, completely crushing the RTX XX90 editions in performance-per-watt, but as soon as memory bandwidth becomes paramount the M-series falls heavily behind.

jayd165• 2 months ago

[..] take full use of the PCIe bandwidth of the GPU or the bandwidth of its GDDR memory.

I'm honestly a little confused by what you mean here. Why would we want to maximize those things? Games are about consistent output under the frame deadline, not full saturation of the hardware.

Why would anyone try to saturate a 5090 with their game? The addressable market is tiny and you'd have to hope their full spec runs as well as or better than your test rig or they'll still not hit framerate.

simonbw• 2 months ago

You could do some sort of adaptive quality where you spend time incrementally improving fidelity until your frame budget is up. In practice I think that might be trickier than it sounds, but I feel like theoretically there's something there that could get you the best graphics your rig can handle without dropping frames. I've been considering doing something like this when I've been building a game/engine lately.

AnthonyMouse4• 2 months ago

Lets systems optimize utilization based on need, rather than be confined to specific pools

The trouble with this is that the different types of memory have different characteristics. Latency for ordinary system memory is actually better than it is for GDDR, because GDDR is optimized for bandwidth. RTX 5090 has 1.8TB/s of memory bandwidth with a 512-bit memory bus. The same bus width for DDR5-9600 would have better latency but only a third of the bandwidth.

CPU workloads are generally bounded by latency and GPU workloads are generally bounded by bandwidth, which is why they use two different types.

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)

The trouble with this is cost. In principle you could get the same 1.8TB/s of memory bandwidth as the RTX 5090 has, with the better latency of DDR5, by using DDR5 with a 1536-bit bus. This is indeed with multi-socket servers do, two sockets with 768-bit in memory channels per socket, but now check how much those system boards cost.

But the remaining alternatives are both worse. If you use GDDR for the unified memory then GDDR costs more than DDR and you're going to have significantly worse latency for the CPU. If you use DDR without a 3-4 times wider bus than the already-wide GPU then the GPU gets starved for bandwidth.

fc417fc8025• 2 months ago

These are all good points that I agree with but rather than seeing an intractable problem I predict we'll see the role that GDDR would otherwise fill in this scenario replaced by a small block of HBM on the APU die. I don't know if it will ultimately end up unified or not but either way I don't think memory segmentation is the core problem here. Simply not needing to send transfers across the narrow and slow PCIe bus would fix most of the practical problems (at least AFAIK but I'm not an expert).

Transitioning over to wild speculation here, I think that most likely this will be treated as part of an absurdly large L3 (ala 3D V-Cache) or as an additional L4. In either case I expect the latency and power tradeoffs introduced to be tolerated as "good enough" even for the highest end consumer gear. (Actually I wonder if some sort of special case cache would be feasible, with memory addresses flagged by the graphics driver and regular CPU related stuff skipping over it entirely. But by then we've squarely entered the territory of vaguely unhinged rambling on my part.)

Alternatively if the performance caveats are deemed to be important enough to justify the added complexity it wouldn't surprise me to see the HBM treated as an independent memory pool analogous to that of a dGPU. That wouldn't change the current status quo with respect to the GPU APIs but it would significantly ameliorate the memory bandwidth bottleneck for inference workloads and from a software perspective is a drop in replacement. You'd still write the code targeting the dGPU with explicit swapping to RAM but when run on an appropriate APU it would get a massive speedup for free instead of suddenly being starved for bandwidth while also performing unnecessary copy operations.

ssivark• 2 months ago

How about having a large pool of unified memory and expanding the next layer (L3?) of cache to accommodate more of the CPU's the low-latency RAM usage?

stego-tech7• 2 months ago

I get all of that already, but stand by my original points: for most consumer, non-data center workloads, the compromises aren’t likely to be noticeable to the end user. We’re not talking about edge cases like local-AI or AAA gaming enthusiasts who want to run software at bleeding-edge capabilities and who will dissect performance deltas between driver versions or overclock their kit for maximum performance, because we’re the edge cases in the marketplace.

Everything is ultimately a compromise of some sort, and modern Unified Memory feels like one of the better compromises out there given the current plateauing of hardware scaling, the growing costs associated with memory and NAND, and the shifting complexity from hardware (more instruction sets, more accelerators, more cores) to software (more abstraction layers, more machine learning).

Melatonic9• 2 months ago

Isn't GDDR also based on a much earlier DDR implementation than DDR5 ?

It also has way better throughput because it's physically surrounding the chip itself and wired in a way that maximises this.

The real problem is interconnect speed and latency. We have made tons of progress elsewhere but AI is exposing that the interconnect in many systems is just not great. Even future PCIE 6.0 is fairly bandwidth constrained compared to 8 channels of DDR memory or the way we solder GDDR next to the chip.

We moved on from AGP and older formats to PCI-E and I think it's time to do that again. And maybe even "slot" based implementations in general for both RAM (system and graphics) and GPUs.

We need consumer and workstations in summary to use pin based stuff like LPCAMM ram. And the interconnect on the motherboard itself needs to be both wider (more bandwidth) and lower latency. This might require moving on from motherboard being 2 dimension only (a flat board) to something like an L shape to gain more physical board space.

wren6991• 2 months ago

Even local AI use cases don’t substantially or meaningfully benefit from faster memory, at least to average consumers.

I'm not sure what you mean by this. Memory bandwidth is the main bottleneck for single-user decode. The bottleneck is actually more severe for end-user inference than cloud inference, because end users don't have the option to increase arithmetic intensity by computing tokens for multiple clients in the same pass.

One thing we've learned from Apple is the viability of spamming more LPDDR5X channels (up to 1024-bit total bus width on M3U) as a means of achieving high bandwidth while keeping the cost/capacity reasonable.

pjjpo• 2 months ago

Isn't the big drawback not having a swappable GPU? Perhaps that's not as important anymore but I'm not sure we've confirmed the market demand for that.

Lplololopo2• 2 months ago

" Even local AI use cases don’t substantially or meaningfully benefit from faster memory, at least to average consumers."

What do you mean by this? Memory bandwidth is fundamental to the speed of an local AI model

mariopt• 2 months ago

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

Gareth3212• 2 months ago

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.

zuzululu• 2 months ago

it should be but im more excited i can game on it

ftchd• 2 months ago

https://xcancel.com/lemire/status/2062880075117113739

proxysna• 2 months ago

It's just a DGX spark with faster memory and a windows boot?

epolanski4• 2 months ago

Not gonna lie, I'm buying one of the 128GB ram ones for local inference if price is human.

throwaway5752• 2 months ago

"Major banana producer suggests shifting more ice cream store menus to banana splits, and increasing the amount of bananas per serving"

Animats2• 2 months ago

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.

forrestthewoods2• 2 months ago

No. It’s all integrated. Not something you can buy separately or upgrade later.

Animats2• 2 months ago

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/

wmf6• 2 months ago

$4,000-5,000

cthalupa7• 2 months ago

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.

GodelNumbering• 2 months ago

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.

moffkalast2• 2 months ago

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.

bigyabai8• 2 months ago

Nvidia can barely scrape together one Ubuntu release per Jetson generation

Why would they do more? It's an LTS distro, the Nvidia drivers are updated for as long as the hardware's compute capability is supported.

Nvidia's ARM drivers are updated constantly, and battle-tested as the backbone in hundreds of thousands of Grace ARM servers.

bigyabai8• 2 months ago

Unlike the M5 Max, it should have usable context prefill. It's feasible to run 256k token workflows that would take the better half of an hour for TTFT on the M5.

xpct• 2 months ago

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.

Someone• 2 months ago

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.”

dh2022• 2 months ago

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…

mrweasel• 2 months ago

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.

kcb• 2 months ago

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.

danielovichdk• 2 months ago

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

dcreater• 2 months ago

They announced RTX spark days ago. Why is this post linking to a "leak" tweet on the frontpage now?

VortexLain• 2 months ago

I really hope this will have proper GNU/Linux support, otherwise it will end up the same way Qualcomm ARM PCs did.

QQ003• 2 months ago

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.

noveltyaccount• 2 months ago

“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.

daft_pink• 2 months ago

Sounds good, but how much does it cost? Is this going to be an affordable laptop or $6000.

shadowpho• 2 months ago

I mean fast 128gb of ram is like $2k so it’ll be $3k after overhead for just the ram portion of device.

htk• 2 months ago

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.

bigyabai8• 2 months ago

The M1 Max has an unusably slow GPU for inference. TTFT on real-world contexts can be over 10 minutes.

Nothing new here, apart from being able to use CUDA on a less power hungry system.

CUDA has been running on ARM SOCs since the Tegra K1, 12 years ago. Nvidia is not new to ARM, nor is CUDA.

einpoklum• 2 months ago

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.

buffer_overlord• 2 months ago

Can it run Ubuntu?

neuroelectron• 2 months ago

It's going to be amazing. Almost twice as fast for only 10 times the heat. Consumers aren't concerned with efficiency they only care about performance.

comandillos• 2 months ago

I dont really get the hype with all the N1X thing when in reality this is the same almost 1 yr old GB10 that was released with the DGX Spark and proved to be quite a disappointment

siliconc0w• 2 months ago

I can't really see wide adoption of local LLMs unless prices really start to climb. It makes sense to use cheaper hosted smaller models like Sonnet or even Kimi but these won't run a Kimi-class model and that is really the floor for non-toy agentic tasks. Spending 5k to avoid a $20 subscription really only makes sense for niche security reasons.

zzzoom• 2 months ago

I'd bet on the inverse: China scaling DRAM production until the price crashes, and the whole US stock market that is propped on top of that scarcity going down with it.

maipen• 2 months ago

Tarrifs.

mathgladiator2• 2 months ago

I think we really haven't even seen how GenAI can influence new products and games. Have you consumed Dungeon Crawler Carl?

satvikpendem2• 2 months ago

Was it made with AI?

vegabook• 2 months ago

I ran two gens of Jetson board and I have zero confidence in this. NVDA is printing in the data centre and everything else has no staying power.

bigyabai8• 2 months ago

I've ran two gens of Jetson, two gens of Nvidia/Linux dGPU and one gen of Tegra. The dGPUs have much better drivers than the devkit hardware does.

mohamedkoubaa• 2 months ago

The commenters here seem to have forgotten that computers can do things other than inference

oldnetguy• 2 months ago

SGI had unified memory back in 1996.

yoyohello13• 2 months ago

If it runs well with Linux, I’m sold. A Windows pc will never see the inside of my network.

seabrookmx• 2 months ago

The DGX spark desktop shipped with Ubuntu, but I haven't seen if it has a bunch of Nvidia specific repos for drivers etc. needed to make it function.

Assuming all that stuff is upstreamed (and they aren't using oddball webcam/input devices etc) it should have much better support than Qualcomm.

Fingers crossed!

Schnitz• 2 months ago

How is this different from something like the AMD Ryzen AI Max that can already be purchased and supports 128GB unified memory? Seriously curious.

nine_k2• 2 months ago

Maybe CUDA support, or something else specific to NVidia?

JBiserkov• 2 months ago

Offtopic, but "Twitter. Now that's a name I've not heard in a long time. A long time".

Aperocky2• 2 months ago

Why is it only for Windows PC, can we not run Linux or at minimum SteamOS?

snvzz• 2 months ago

It is not RISC-V.

We aren't so naive as to move from a locked IP ISA like x86 to another locked IP ISA such as ARM.

Right?

gghh2• 2 months ago

Sorry for lazy question but some people here know this off top of their head so asking. Memory bandwidth of this chip?

Last time I check an NVidia situation was for DGX Spark (the GB10 chip), it has regular LPDDR5X which by JEDEC standard cannot go beyond ~270 GB/sec, ie 8533 Mbit/s on a 256 lanes bus.

So yeah Lemire seems to go "OMG unified memory, they're following Apple path..." ok, but Apple pulled off a much faster interconnect, 800 GB/s ballpark, and I'm trying to understand (not really, I'm asking you to try understand, he he) how is this laptop faring in that regard.

easygenes3• 2 months ago

This is the same chip and same memory. Only difference is it is going in a laptop, so will be more thermally limited.

gghh2• 2 months ago

Gotcha, thanks for explaining. Indeed I noticed it was 128G of memory just like the DGX.

1970-01-01• 2 months ago

Local models becoming thousands of dollars instead of millions to run is a story the public genuinely seems to be unaware of. If the order of magnitude falls again, the markets are cooked. The cheap chips barrier is even artificial and unsustainable. The next big story in local AI adoption will be big players doing chip hoarding-as-a-strategy.

wombat-man• 2 months ago

I think this is part of why capex is so high for the big players. Yes data centers are expensive to build but it's also about outbidding for the labor and materials needed to build those. They'd rather build compute they don't need and then lease it then let competitors have their own machines.

iceflinger• 2 months ago

I'm fully convinced now that this is the real reason for the OpenAI RAM wafer deals.

effnorwood• 2 months ago

beast is right

Melatonic9• 2 months ago

We need a physical architectural change if we really want to improve this. You can only physically wire so much memory next to a chip in a 2 dimensional (flat) design - limited by the edge lengths of the chip (CPU or GPU)

Basically what we need is a chip that also has pins or some type of attachement system on the top (physically) or maybe below where the chip itself connects to the motherboard.

Imagine a CPU you can just plug in a block of HBM memory on top of (or on "bottom" of). This would allow a much larger physical surface area for putting ram cache near the compute cores itself because you would not be limited to edge lengths.

Cooling the whole thing would be a methodology change (might need liquid coolers that sandwhich in between the ram cache and compute and cool both)

Melatonic9• 2 months ago

What does he mean at the end when he talks about Intel processors - is he referencing something supported in AVX-512 that is unique to AMD or ?

empiree• 2 months ago

Interesting move. More competition in CPUs is good for everyone. My only question is software support, that's usually where these non-x86 launches stall. Hoping it's more than a spec-sheet flex.

mindcrash• 1 month ago

"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"

Not really a game changer as AMD already has this (but is quite terrible at marketing it). This is the reason why you can run pretty big AI models on reasonable speeds on embedded GPUs on the Ryzen AI SoCs and get even better performance on the Ryzen AI Max+ SoCs.

As a example I have a HX370 based laptop with 128G RAM and due to these specs can instruct the included firmware to share up to 96G (or approximately 66% of available system memory) with the embedded GPU, which gives quite some headroom to experiment with local models as said before.

The included GPU with 6144 CUDA cores isn't really big news either to be honest, as the DGX Spark already has them too. Likely because they simply integrated the same (or very similar) GPU on this SoC. In fact the whole thing looks eerily similar to the DGX Spark SoC, so the only new thing pretty much seems to be the support for Windows.

Nvidia is proposing a beast of a CPU system for Windows PCs

Related Stories