It was an open secret that NVIDIA was working on an ARM-based system-on-a-chip for laptops and desktops, and today at Computex 2026 the company unveiled what it’s been working on. It’s surely a beast, and unsurprisingly, it’s lathered in “AI” buzzwords.
At full strength, this chip offers up to 20 Arm CPU cores, a Blackwell GPU with 6,144 CUDA cores, 128GB of LPDDR5X RAM, and up to 300 GB/s of memory bandwidth. That powerful CPU and GPU, connected over NVLink C2C, and the large memory pool give AI agents and 120-billion-parameter models plenty of power and space for long-running tasks with context lengths stretching to a million tokens, according to Nvidia.
RTX Spark will power high-end laptops from partners including Dell, HP, Lenovo, Asus, and MSI — and notably, a new Surface Ultra laptop from Microsoft. Nvidia says it’s worked with those partners to create “the most extraordinary laptops [they’ve] ever built,” with tandem OLED G-Sync displays, “all-day” battery life, premium aluminum chassis with large glass touchpads.
↫ Jeffrey Kampman at Tom’s Hardware
I couldn’t care less about the “AI” nonsense, but the chip itself seems like an absolute monster for laptops and mini PCs. With that much power and a solid NVIDIA GPU, these are also great for gaming and creative tasks, making them feel like the first true competition in the PC space to Apple’s M series of chips. They’re planned for late 2026, and tellingly, there’s no pricing information just yet.
Hopefully Linux works well on them and is not left out.
If they are Windows only that’s still a flop as Windows for ARM sucks. (I am also reading that benchmarks from last year put them about 2 years behind Apple)
It’s Nvidia, so my hopes are low. Even when they have to support Desktop Linux (on DGX Spark for example), they do it using proprietary distros with proprietary drivers.
Generally, ARM “PCs” with non-existent official support for Desktop Linux (think Apple Silicon) and/or Microsoft locking down Secure Boot on actual PCs (mandating it’s always on and boot only “approved” OSes) is a future Linux people haven’t prepared for.
I already experienced the second to a degree: on new laptops, you have to go to the UEFI settings and enable “3rd Party UEFI CA” to boot a Desktop Linux Live USB, I had to set up a Lenovo laptop for work, and came across this.
kurkosdr,
Getting nvidia GPU drivers working on ARM isn’t unheard of…
https://alican-kiraz1.medium.com/run-llm-on-pi5-connecting-an-nvidia-gpu-to-raspberry-pi-5-via-pcie-x4-a6d52c3efd2a
But I agree it’s disappointing that they didn’t say anything about linux support.
I’m not thrilled about nvidia continuing to use proprietary drivers. Still, ever since nvidia mainlined their own stable driver interface into linux a couple of years back, using nvidia’s proprietary drivers has become a lot more robust than their old process. I haven’t experienced any failures since nvidia’s stable ABI got mainlined even with unsupported operating systems. Mind you this was with x11, I’ve heard wayland users were complaining more, but most of that friction is probably behind us.
There’s an ideological gap. Many linux users, myself included, want manufacturers to release source code, but over the decades this hasn’t panned out and we’re always stuck having to reverse engineer everything. Using the manufacturer’s proprietary drivers means we’re stuck with the kernels blessed by manufacturers and it’s negatively impacted android and linux SBC use cases for decades. The alternative is to support a stable ABI so that manufacturer drivers can work across numerous kernels, even after updating to a new OS. Alas, we’re in an ideological stalemate and end users have been left suffering for it. Nvidia were rather lucky they got a stable userspace ABI into linux for their proprietary drivers.
I’m convinced that there are many people who can solve the technical problems, but the ideological impasse between open drivers and stable ABIs might never be resolved.
it could be a while before I see this because I think it’s a new problem and I usually buy used computers, but I agree these kinds of barriers can happen.
Couldn’t agree more about the ideology impasse.
Hardware companies don’t always want to release their code, especially driver code if that is part of what gives them a competitive advantage.
So even if they Do release Linux compatibility, if it’s not open source, it gets shunned by the “community”
It’s not ideology, it’s the architecture and Linux kernel development process. Drivers have always been part of the kernel and their maintenance is integral part of kernel development process.
Being able to break ABI is why Linux is portable and can iterate quickly.
If maintaining ABI was that simple a company with enough financial muscle would step up and maintain an up to date fork of Linux with stable driver ABI (time limited of course). So far it haven’t happened. It haven’t even after GPU drivers became as important as RAID drivers in the datacenter.
dsmogor,
This unstable approach “let us never commit to a stable kernel ABI or API” *is* ideology though. And there are real pros and cons. Driver incompatibility across kernels is obviously a con. People might assume it’s just proprietary drivers. but even when you have the source code and have to maintain it, the unstable ABI is huge source of churn. I’ve sat at that seat and I hated how much overhead unstable ABIs create for development. Unfortunately this discourages manufacturers from committing to linux support.
Moreover If you depend on multiple out of tree drivers, you can end up with one driver supporting one version of the kernel and another driver supporting another version of the kernel and making them all compatible is left as an exercise for downstream distros/users.
Well, the problem isn’t that there aren’t developers who could do it, it’s that most of them don’t get to decide what patches get mainlined. Even google themselves have fought with mainline at times. The problems are fixable. Even without long term stability, it would be reasonable to have stability at least within major kernel versions. This solves a lot of our problems while also alleviating the concerns over long term ABI support,
It’s not a pie in the sky idea, many operating systems have ABI stability, including linux when they want to. We could do it if not for the significant ideological barriers in play.
It did happen for nvidia, who got their stable ABI into linux and it fixed practically all the driver deployment problems that plagued nvidia drivers on linux. Now the nvidia drivers are legitimately compatible against a wide range of kernels. Granted being a userspace ABI helped sidestep the opposition to having stable ABIs in the kernel, but regardless it demonstrates how stable ABIs can solve linux driver issues and are a boon for alternative distros and users. I’d love for this to be the case across all android and linux SBCs devices.
The main detractor isn’t technical at all, but the fact that stable ABIs enable manufacturers to release proprietary drivers instead of open source code. I agree proprietary drivers sucks, however they’ve already been depriving us of open source for decades anyway. Over all this time we haven’t won that battle and we probably never will. Meanwhile stable ABIs fix so many of the driver compatibility problems, I honestly don’t think are ever going to get fixed otherwise.
> It’s Nvidia, so my hopes are low
The primary operating system for the DGX Spark was Linux (Ubuntu). This is essentially the same hardware.
https://nvdam.widen.net/s/tlzm8smqjx/workstation-datasheet-dgx-spark-gtc25-spring-nvidia-us-3716899-web
In Nvidia-land, there is a huge difference between "essentially the same hardware" and "the same hardware".
Want some examples?
– The Area-51m laptop has an RTX 2080 GPU that's "essentially the same" as desktop RTX 2080 GPUs, but can't use the WHQL drivers without Inf-hacking, only the comparatively crappier DCH drivers
– Similarly, I had to Inf hack the drivers to make the Dell-ROM 770M GPU in an Alienware 17 R1 laptop work with the latest Nvidia drivers (they fixed that in the latest drivers, but my point stands).
– RTX 2080 SUPER GPUs can't use Nvidia 3D Vision without 3rd party hacks, despite being "essentially the same hardware" as RTX 2080 non-SUPER GPUs
Remember, this is all proprietary drivers, you can't just take the source code and modify it to work with "essentially the same hardware".
This is the same chip basically as the one in the DGX Spark, so Linux support should be fairly straightforward.
Windows on ARM is fine nowadays.
Pretty sure gaming on ARM/Windows is still in a weird state. Errors like processor extensions not detected, bugs like needing to touch the touchscreen to stop the game camera from freaking out. OpenGL being implemented by the GPU manufacturer. ARM laptops with Nvidia/AMD gpus should be more stable/compatible, but this looks like another $4000 AI dev machine.
See my comment above, in Nvidia-land, there is a huge difference between “basically the same chip” and “the same chip”.
Bro, you mentioned getting a different driver to work via inf-hacking. That’s evidence that it _is_ basically the same hardware, not evidence that it’s not. If it weren’t, you’d have to do a lot more than change an ID in an inf
Reading comprehension, can we get some reading comprehension over here?
@kurkosdr literally said that it is essentially/basically the same hardware. Like, those are the exact words he used. However, Nvidia purposely changed it just enough that the same driver can’t work on both devices without hacks. If it’s different enough to **require** a hack, it’s different hardware, period, from Nvidia’s point of view. And I agree with him, it amounts to a huge difference, especially to your average everyday user who doesn’t even know what an INF file is, they just want their $3000 gaming laptop to stop shitting the bed in every game they try to play compared to their perfectly working desktop with “basically the same hardware”.
That’s why there is concern that the DGX and RTX might not work the same in Linux, or that the latter might not work at all in Linux, because Nvidia controls the hardware as well as the proprietary drivers and are known for fuckery around drivers. Don’t be surprised if running Linux on the RTX Spark requires a lot more than a driver definition hack, as it’s clear this is aimed at Windows support first.
The DCH drivers are acceptable for regular 2D, it’s Nvidia 3D Vision where they poop the bed. And that, despite the fact that Nvidia 3D Vision was (theoretically) supported by Nvidia at the time those DCH drivers were released.
As a further aside, after I installed the inf-hacked WHQL drivers, the GPU of the Area-51m showed up as the “Max-Q” variant of RTX 2080 in Device Manager, despite the fact I put it in the Inf file as a plain (non-Max-Q) RTX 2080 (because that’s what the laptop has). Lord knows why did that happen, but it did. And hacking the inf files isn’t as simple as it sounds, there are multiple sections for multiple variants of the “RTX 2080” spread across multiple inf files.
The reason I am saying this is that, if you want to do something that’s is officially unsupported, such as installing Desktop Linux on an ARM box like the RTX Spark that doesn’t officially support Desktop Linux (or for the case of Nvidia 3D Vision, something that is theoretically supported but in practice abandoned by Nvidia since 2014 and left to bit-rot), you can’t rely on the whole “essentially the same hardware” belief, not when it comes to Nvidia’s proprietary drivers.
Yes, we all saw the same comment you keep spamming.
Sorry, nothing annoys me more than nerds assuming things “should” work because “it’s essentially the same hardware”. No, that’s not how proprietary drivers (and firmware) works.
kurkosdr,
That’s true, even though it’s supposed to be essentially the same technology, we can’t really know if there are other driver barriers in place. And because of nvidia’s proprietary drivers linux devs may not be able to independently fix them.
I think we all agree that it would be really trivial for nvidia to add linux support. Nvidia have shown they can support linux when they want to, the customers are there, and it works decently.. But that’s what makes it all the more alarming that they’re not officially doing so here. Who knows, they may actually have a deal with microsoft to make it a windows exclusive?
I’m sure people will be quick to try and install linux when the product is out and then all our questions will be answered.
> Hopefully Linux works well on them
This is the same hardware that shipped with the DGX Sparc and it ran Ubuntu out-of-the-box. So, I am going to say that this thing CAN run Linux.
> and is not left out
It has been left out. This is being announced as a partnership with Microsoft and it runs Windows 11 natively.
The question is if they will actively prevent you from running Linux on it. Somebody has to buy one to find that out.
No, it’s not the same hardware, and in Nvidia-land, there is a huge difference between “basically the same hardware” and “the same hardware”.
Don’t hold it against me when I say I will not hold my breath. First it is NVIDIA. NVIDIA and Linux just isn’t the ideal pairing. Second it is ARM. 100% potential and maybe 25% realization? God knows what badly supported system images this beast will need, if Linux is even considered. ARM doesn’t have a standardized ecosystem and therefore there are as many ARM platforms as there are device models in the wild.
Since we are in the middle of the RAMpocalypse, introducing a new high-end platform with 128GB RAM is practically DOA. The cost will either be exorbitant or the machine will be crippled with 8GB RAM. It’s not only system memory that is in a massive crunch, the SSDs are experiencing a price hike too. I can’t see a successful launch under these conditions and that is disregarding the problems with NVIDIA and ARM in general.
I don’t know if this is on purpose, but on-die memory (presumably with a bus shared by the CPU and the GPU cores) might end up being very useful for local LLM inference, in the same way the Mac Mini is the darling of that weird OpenClaw crowd… I know it’s deeply problematic, but IMO at least a few of the moral issues surrounding LLMs (data centre construction, privacy risks, high billing costs) are resolved by local inference.
LLMs aren’t entirely useless… honestly I’ve been dragged kicking and screaming into the world of AI-assisted development, and I’m still screaming a fair bit. Still not convinced it’s worth its salt, but I’m trying to keep an open mind.
skeezix,
High end consumer GPUs are faster than apple’s offerings on compute, but as I’m sure you know, are memory constrained and therefor can’t run larger models. Apple M chips suffer iGPU bottlenecks wherein CPU and GPU cores are forced to compete for memory/power/thermals/etc, however it is actually an excellent architecture for local LLM inference because of how much ram you can get and the con of iGPUs aren’t as applicable to LLMs. I don’t have many applications for apple hardware, but I would agree they are very compelling for local consumer LLMs.
I’m not sold on them everywhere, but I do think they’ll become increasingly used for automating tasks and interacting with humans. Stores are itching to replace all the human labor which machines and TBH this actually bothers me a great deal. I fight these trends as much as I can with my wallet, but…. yeah 10-15 years from now most of those jobs may be gone as interacting with machines continues to become normalized.
This is exactly what I thought when I read this news. They are trying to compete with Apple’s unified memory architecture for people who want to run their LLMs in their own machines. Currently Mac minis are awesome solutions for anyone who is content with a 27B or 31B model. Heck, some ultra machines from Apple can even run things comparable to what openai or anthropic offered last year. Those are currently beyond the reaches of x86 machines with even the most expensive rtx cards.
cevvalkoala,
Are you including RTX PRO 6000 with 96GB VRAM? Keep in mind this is dedicated VRAM in addition to system ram whereas a mac computer needs to share that RAM with the OS/caching/CPU tasks.
RTX hardware is just a lot faster than apple’s M5…
https://www.hardware-corner.net/m5-max-local-llm-benchmarks-20261233/
However, nvidia GPUs come at a price…
Given the substantial performance difference, the utility of CUDA over Metal for me, and usability under linux, I’d rather an RTX pro 6000 if money weren’t an issue. Given the price though the RTX pro is just not affordable. An M5 mac can get you in the game while saving $4-5k versus what it would cost to build a full RTX pro 6000 computer. If you can work with it’s 32GB VRAM, an RTX5900 build would delivers better performance at a lower cost than the mac M5. I agree with this conclusion:
It will be interesting to see where RTX Spark lands in terms of both performance and cost. And since personally I have no interest if it doesn’t run linux. Apple’s M3/M4/M5 macs can’t run linux either. Who knows how long it will be before the Asahi project gets linux running on these architecture AND have full GPU acceleration.
https://asahilinux.org/docs/platform/feature-support/overview/
RTX spark could get there first.
Yeah, you’re right on both accounts. Theoretically there are Nvidia cards that are faster than apple hardware. And as you pointed out, they are ridiculously priced. I should have made a more accurate statement.
That’s not true, stacking up 4 4090s will give you better performance (not per watt) than Mac offerings for ~the same HW investment.
But it’s more difficult to maintain.
That is not on-die memory, it’s on-package memory.
This chip is derived from the current Blackwell SoC in the DGX Spark, so it was literally originally intended for AI workloads.
oh hm, that’s too bad. I read something to that effect yesterday too — the DGX spark, the machine that begat this chip, has not-that-great memory performance. Oh well.
BG10 is at M4 Pro or Strix Halo level and of memory BW, so if they keep similar number of mem controllers for the mobile SKUs, that’s should be pretty good for a mobile SoC.
I’ll wait for benchmarks before I assume that this chip is an absolute beast. I REALLY don’t like that they’re talking “cores” when they talk performance.
Just for reference, an RTX 3080 or 3090 (c.2020) has roughly the same number of CUDA cores as this chip, and 20 ARM Cores isn’t necessarily anything to brag about if their ARM cores are anything like e.g. Qualcomm’s (i.e. roughly half as performant per core as Apple’s M-series chips).
They’re also talking about “1440p gaming at 100 FPS” which is an odd brag. That sounds like they can’t claim a frame rate people actually use (120-144), which if true means that — in the real world — you’re just doing 60 FPS.
From a product standpoint, this all also depends on Microsoft and Adobe delivering performance optimizations tailored specifically for this chip. Which on paper they might, but also they didn’t bother for e.g. Qualcomm with the X2.
Brainworm
A blackwell GPU with 128GB ram sounds like it could be really good on paper, but yeah we don’t know how it will actually perform until 3rd parties benchmark it. I don’t put much faith in self-reported benchmarks because too often it doesn’t reflect the engineers’ real findings after going through the marketing BS department: cherry-picking, misleading claims, incomplete data, unfair comparisons, bad graphs, etc. While some manufacturers have been particularly notorious for these schemes (ahem Apple), all manufacturers need to be scrutinized.
Also, this point is more about the state of the market than anything technical, but if this chip performs as well as an RTX 6000 pro retailing for $8-10k and has more memory, then I don’t see how this product could be significantly more affordable and stay in stock. So although “game ready” implies consumer use cases, I have significant doubts over average consumers being able to get their hands on it. These are probably going to end up in the hands of companies & professionals creating generative content. To this end the Adobe tie-ins make a lot of sense – people will pay for that.
I mean, the chip is basically a BG10 derivative, so we sort of know what the upper performance to expect benchmark wise.
Xanady Asem,
You’ve said BG10 a couple times although but I think you mean GB10.
I’ve seen multiple outlets put it’s performance around a 3080/5060..
https://www.techpowerup.com/gpu-specs/gb10.c4342
Assuming the RTX spark’s performance ends up being similar, it will measure up well as an iGPU, but lackluster compared to high end discrete GPUs. However it still seems to me that these could go into high demand over the amount of ram they offer to handle large models.
RTX SPARC chip? Oh no, got all excited it might breathe new life into Sun Micro’s old ISA, but turns out it is just another ARM implementation. Here’s hoping it is fully accessible to open source OS such as Genode/Sculpt.
There is only Intel/AMD, ARM, and RISC-V now. It is unlikely we will see another mainstream ISA in my lifetime.
Realistically, only two of the three above are likely to remain commercially viable in 15 years from now.
It’s RTX Spark, not RTX Sparc. Also, SPARC sucked anyway, the “register windows” feature forces hardware architects to waste registers for a feature that yields bugger-all in terms of performance. And of course, the feature is mandatory.
Don’t get me wrong, I like Sun Ultra 25/45 and Sun Blade 1500/2500 workstations (and the Solaris OS they run) just as much as the every nerdy person who became acquainted with them in university, but this doesn’t say anything about SPARC as an ISA.
As ever this will live or die by the same players.
Photoshop, CAD and Games (steam)
Can Windows run them at equivalent or better performance to get enough market share to justify Photoshop targeting it?
The alternative is Nvidia throw a few billion at them to make it happen….
Intel and AMD need to decide what the future looks like. ARM is here to stay as a real player.
Do you think Photoshop will remain a source of attraction for years to come?
I mean, just like LLMs took away my job (translation), they are likely to take away graphic designer jobs, especially entry level ones. Install ComfyUI and tell the thing to do whatever you like on the image. That’s getting easier and easier. The market for Photoshop will not remain as big as it was, I reckon. What is your take on this?
Based on current ram prices these will only cost infinity money.
Really? How does an individual of those 20 cores stack up to Apple M4? The sheer number of them says that there’s not as much silicon dedicated to an individual core.
How will conventional LPDDR do sharing bandwidth between the CPU and GPU compared to HBM?
Indeed for dedicated AI pipelines where CPU is just making sure data is loaded from HDD this setup might be promising alternative to a Mac (but where’s 256 and 512 MB version).
These cores appear to be the same as the ones used in other Blackwell SKUs, so CPU performance should be roughly in M2 territory, about two generations behind Apple’s M5.
The memory architecture has also effectively been measured already, since these are, for all intents and purposes, derivatives of the GB10 used in DGX Spark. That puts them around M4 Pro territory in terms of memory bandwidth. Nowhere near HBM, but then again HBM kind of sucks for the type of power/thermal envelopes expected in laptops.
Nvidia is kind of OS/2ing themselves on desktop LLMs because every user that gets enough compute to run decent models locally stops renting so much time from Nvidia’s biggest datacenter customers. They won’t ship a 1TB RAM Spark unless and until Starship+Terafab work as advertised and renting sat time from Elon gets more cost effective than building GPU datacenters with brutal capex depreciation.