From the FAQ… doesn’t seem promising when they ask and then evade a crucial question.
> What is the memory bandwidth supported by Ascent GX10? AI applications often require a bigger memory. With the NVIDIA Blackwell GPU that supports 128GB of unified memory, ASUS Ascent GX10 is an AI supercomputer that enables faster training, better real-time inference, and support larger models like LLMs.
LLM performance depends on doing a lot of math on a lot of different numbers. For example, if your model has 8 billion parameters, and each parameter is one byte, then for 256gb/s you can't do better than 32 tokens per second. So if you try to load a model that's 80 gigs, you only get 3.2 tokens per second, which is kinda bad for something that costs 3-4k.
There's newer models called "Mixture of Experts" that are, say, 120b parameters, but only use 5b parameters per token (the specific parameters are chosen via a much smaller routing model). That is the kind of model that excels on this machine. Unfortunately again, those models work really well when doing hybrid inference, because the GPU can handle the small-but-computationally-complex fully connected layers while the CPU can handle the large-but-computationally-easy expert layers.
This product doesn't really have a niche for inference. For training and prototyping is another story, but I'm a noob on those topics.
Running LLMs will be slow and training them is basically out of the question. You can get a Framework Desktop with similar bandwidth for less than a third of the price of this thing (though that isn't NVIDIA).
They have failed to provide answers to other FAQ as well. The answers are really awkward and don't read like LLM output which I'd expect to be much more fluent. Perhaps a model which was lobotomized through FP4 quantisation and "fine tuning" on one of these.
For comparison, the RTX 5090 has a memory bandwidth of 1,792 GB/s. The GX10 will likely be quite disappointing in terms of tokens per second and therefore not well suited for real-time interaction with a state-of-the-art large language model or as a coding assistant.
Seems this is basically DGX Spark with 1TB of disk so about $1000 bucks cheaper. DGX Spark has not been received well (at least online, Carmack saying it runs at half the spec, low memory bandwidth etc.) so perhaps this is way to reduce buyers regret, you are out only $3000 and not $4000 (with DGX Spark).
He is very enthusiastic about new things but even he struggled (for ex. the first link is about his experience OOB with Sparq and it wasn't a smashing success).
Should you get one? #
It’s a bit too early for me to provide a confident recommendation concerning this machine. As indicated above, I’ve had a tough time figuring out how best to put it to use, largely through my own inexperience with CUDA, ARM64 and Ubuntu GPU machines in general.
The ecosystem improvements in just the past 24 hours have been very reassuring though. I expect it will be clear within a few weeks how well supported this machine is going to be.
Don't undersell it. The game is playable in a browser. The graphics are just blocks, the aliens don't return fire. There are no bunkers. The aliens change colors when they descend to a new level (whoops). But for less than 60 seconds of effort it does include the aliens (who do properly go all the way to the edges, so the strategy of shooting the sides off of the formation still works--not every implementation gets that part right), and it does detect when you have won the game. The tank and the bullets work, and it even maintains the limit on the number of bullets you can have in the air at once. However, the bullets are not destroyed by the aliens so a single shot can wipe out half of a column. It also doesn't have the formation speed up as you destroy the aliens.
So it is severely underbaked but the base gameplay is there. Roughly what you would expect out of a LLM given only the high level objective. I would expect an hour or so of vibe coding would probably result in something reasonably complete before you started bumping up into the context window. I'm honestly kind of impressed that it worked at all given the minuscule amount of human input that went into that prompt.
I do think that people typically undersell the ability of LLMs as coding assistants!
I'm not quite sure how impressed to be by the LLM's output here. Surely there are quite a few simple Space Invaders implementations that made it into the training corpus. So the amount of work the LLM did here may have been relatively small; more of a simple regurgitation?
>The aliens change colors when they descend to a new level (whoops).
That is how Space Invaders originally worked, used strips of colored cellophane to give the B&W graphics color and the aliens moved behind a different colored strip on each level down. So, maybe not an whoops?
Edit: After some reading, I guess it was the second release of Space Invaders which had the aliens change color as they dropped, first version only used the cellophane for a couple parts of the screen.
Some of the stuff in the Carmack thread made it sound like it could be due to thermals, so maybe could reach or come a lot closer to, but not sustain, and if this has better cooling maybe it does better? I might be off on that.
I don't understand DGX Spark hate. It's clearly not about performance (a small, low-TDP device), but ability to experiment with bigger models. I.e. a niche between 5090 and 6000 Pro, and specifically for people who want CUDA
Wasn't it shown that Carmack just had incorrect expectations, based upon misunderstanding the details of the GPU hardware?
From rough memory, something along the lines of "it's an RTX, not RTX Pro class of GPU" so the core layout is different from what he was basing his initial expectations upon.
Except Carmack, as much as I hate to say it, was simply wrong. If you run the GPU at full throttle then you get the power draw that he reported. However, if you run the CPU AND the GPU at full throttle, then you can draw all the power that’s available.
This is a tangent, but the little pop up example for their ai chat bot to try and entice me to use it was something along the lines of “what are the specs?”
How great would it be if instead of shoving these bots to help decipher the marketing speak they just had the specs right up front?
I find all these Popup Assistant Bots as bad User Experience.
No, I don't want to use your assistant and your are forcing me to pointlessly click on the close button. Some times they event hide viable information during their popup.
They seem to be the reincarnation of 2000s popups; there to satisfy a business manager versus actually being a useful tool.
I had one of these on pre-order/reservation from when they announced the DGX Spark and ended up returning it after a couple days. I thought I'd give it a shot, though. The 128GB of unified memory was the big selling point (as are any of the DGX Spark boxes), but the memory bandwidth was very disappointing. Being able to load a 100B+ parameter model was cool in terms of novelty but not particularly great for local inferencing.
Also, NVIDIA's software they have you install on another machine to use it is garbage. They tried to make it sort of appliance-y but most people would rather just have SSH work out of the box and can go from there. IMO just totally unnecessary. The software aspect was what put me over the edge.
Maybe the gen 2 will be better, but unless you have a really specific use case that this solves well, buy credits or something somewhere else.
If (and in case of Nvidia that's a big if at the moment) they get their software straight on Linux for once this piece of hardware seems to be something to keep an eye on.
Test Model Metric EVO – X2 NVIDIA GB10 Winner
Llama 3.3 70B Generation Speed (tok/sec) 4.90 4.67 AMD
First Token Response Time (s) 0.86 0.53 NVIDIA
Qwen3 Coder Generation Speed (tok/sec) 35.13 38.03 NVIDIA
First Token Response Time (s) 0.13 0.42 AMD
GPT-OSS 20B Generation Speed (tok/sec) 64.69 60.33 AMD
First Token Response Time (s) 0.19 0.44 AMD
Qwen3 0.6B Model Generation Speed (tok/sec) 163.78 174.29 NVIDIA
First Token Response Time (s) 0.02 0.03 AMD
And additionally Framework apparently benchmarked GPT-OSS 120B (!) on the maxed out 395+ Desktop and reached a 38.0 tok/sec Generation Speed. Given that Nvidia can't even keep up on a 20B model, I assume they can't keep up on the 120B model aswell.
So to me the only thing which seems to be interesting about the Spark atm is the ability to daisy link several units together so you can create a InfiniBand-ish network at InfiniBand speeds of Sparks.
But overall for just plain development and experimentation, and since I don't work at Big AI, I'm pretty sure I would not purchase Nvidia at the moment.
Unfortunately comparing tok/sec right now in a vacuum and especially across weeks of time is kind of pointless. Everything is still evolving; there were patches within days that bumped GB10 performance by double digit percentiles in some frameworks. You just kind of have to accept things are a moving target.
For comparison, as of right now, I can run GPT-OSS 120b @ 59 tok/sec, using llama.cpp (revision 395e286bc) and Unsloth dynamic 4-bit quantized models.[1] GPT-OSS 20b @ 88 tok/sec [2]. The MXFP4 variant comes in the same, at ~89 tok/sec[3]. It's probably faster on other frameworks, llama.cpp is known to not be the fastest. I don't know what LM Studio backend they used. All of these numbers put the GB10 well ahead of Strix Halo, if only going by the numbers we see here.
If the AMD software wasn't also comparatively optimized by the same amount in the same timeframe, then the GB10 would be faster, now. Maybe it was optimized just as much; I don't have a Strix Halo part to compare. But my point is, don't just compare numbers from two various points in time, it's going to be very misleading.
These are valid points but the numbers are still useful as a floor on performance.
Given Strix Halo is so much cheaper I'd expect more people to work on improving it, but the NVIDIA tools are better so unclear which has more headroom.
Yeah that's fair. 60 tok/sec on a gpt-oss-120b is certainly nice to know if you should even think about it at all. I'm quite happy with it anyway.
The pricing is definitely by far the worst part of all of this. I suspect the GB10 still has more perf left on the table, Blackwell has been a rough launch. But I'm not sure it's $2000 better if you're just looking to get a fun little AI machine to do embeddings/vision/LLMs on?
> If (and in case of Nvidia that's a big if at the moment) they get their software straight on Linux for
What exactly isn't working for you? The last two/three months I've been almost exclusively doing ML work (+CUDA) with a NVIDIA card on Linux, and everything seems to work out of the box, including debugging/introspection tools and everything else I've tried. As an extra plus, everything runs much faster on Linux than the very same hardware and software does on Windows.
On the networking side. M4 max does have thunderbolt 5, 80gbps advertised.
Would ip over TB not allow for significantly faster interconnects when clustering Macs?
The DGX Spark, Ascent GX10, and related machines have no relation to NVIDIA Grace Blackwell GB200. The chip they are based on is called GB10, and is architecturally very different from NVIDIA's datacenter solutions, in addition to being vastly smaller and less powerful. They don't have anything resembling the Grace CPU NVIDIA used in Grace Hopper and Grace Blackwell datacenter products. The CPU portion of GB10 is a Mediatek phone chip's CPU complex that metastasized, not NVIDIA's datacenter CPU cut down.
Mediatek is in the picture because NVIDIA outsourced everything in GB10 but the GPU to Mediatek. GB10 is two chiplets, and the larger one is from Mediatek. Yes, Mediatek uses off the shelf ARM CPU core IP, but they still had to make a lot of decisions about how to implement it: how many cores, what cluster and cache arrangements, none of which resemble NVIDIA's Grace CPU.
Thanks for the clarification. I was surprised to learn it is not a single chip; thought they did something akin to Apple Silicon integrating some ARM cores on their main chip. Kind of disappointing: they basically asked MediaTek to build a CPU with an NV-Link I/O.
The big picture is probably that GB10 is destined to show up in laptops, but NVIDIA couldn't be bothered to do all the boring work of building the rest of the SoC and Mediatek was the cheapest and easiest partner available. It'll eventually be followed by an Intel SoC with NVIDIA providing the GPU chiplet, but in the meantime the Mediatek CPU solution is good enough.
From NVIDIA's perspective, they need an answer to the growing segment of SoCs with decent sized GPUs and unified memory; their existing solutions at the far end of a PCIe link with a small pool of their own memory just can't work for some important use cases, and providing GPU chiplets to be integrated into other SoCs is how they avoid losing ground in these markets without the expense of building their own full consumer hardware platform and going to war with all of Apple, Intel, AMD, Qualcomm.
I wonder why they even added this to the FAQ if they're gonna weasel their way around it and not answer properly?
> What is the memory bandwidth supported by Ascent GX10?
> AI applications often require a bigger memory. With the NVIDIA Blackwell GPU that supports 128GB of unified memory, ASUS Ascent GX10 is an AI supercomputer that enables faster training, better real-time inference, and support larger models like LLMs.
Never seen anything like that before. I wonder if this product page is actually done and was ready to be public?
Taiwanese companies are legendary for producing baller hardware with terrible marketing and documentation that answers important questions. It's like those teams don't talk to each other inside the business.
Fortunately, their products are also easy to crack open and probe.
It seamlessly combines Nvidia's price gouging and ASUS's shady tactics. God forbid you ever have to RMA it, they'll probably brake it and blame it on you.
How is it different from their consumer GPU marketing? They have Founder Edition under NVIDIA brand initially, but the ecosystem is supposed to mass produce. It appears to be the same for DGX Spark where PNY has produced the NVIDIA branded and now you're going to see ASUS and Dell and others make similar PCs under their brand.
As far as I can tell these are all the same hardware just different enclosures. I'm not sure why Nvidia went this route given that they have a first party device. Usually you only see this when the original manufacturer doesn't want to be in the distribution or support game.
If this is anything like their consumer graphics cards, the first-party version will only be available in the dozen or so countries where Nvidia has established direct distribution channels and they'll defer to the third parties everywhere else.
It's just Ubuntu with precanned Nvidia software, otherwise it's a "normal" UEFI + ACPI booting machine, just like any x86 desktop. People have already installed NixOS and Fedora 43, and you can even go ahead and then install CUDA and it will work, too. (You might be able to forgo the nvidia modules and run upstream Mesa+NVK, even.) It's very different from Jetson and much more like a normal x86 desktop.
The kernel is patched (and maintained by Canonical, not Nvidia) but the patches hanging off their 6.17-next branch didn't look outrageous to me. The main hitch right now is that upstream doesn't have a Realtek r8127 driver for the ethernet controller. There were also some mediatek-related patches that were probably relevant as they designed the CPU die.
Overall it feels close to full upstream support (to be clear: you CAN boot this system with a fully upstream kernel, today). And booting with UEFI means you can just use the nvidia patches on $YOUR_FAVORITE_DISTRO and reboot, no need to fiddle with or inject the proper device trees or whatever.
That was also my experience with their Jetson series [1], but my understanding is that these DGX kernels are not maintained by Nvidia but by Canonical, so they operate directly out of their package repos and on Canonicals' release and support schedule (e.g. 24.04 supported until 2029.) You can already get 6.14 from the package repos, and 6.17 can be built from source and is regularly updated if you follow the Git repositories. It's also not like the system is unusable without patches, and I suspect most will go upstream.
Based on my experience it feels quite different and much closer to a normal x86 machine, probably intentional. Maybe it helped that Nvidia did not design the full CPU complex, Mediatek did that.
[1] They even claim that Thor is now fully SBSA compliant (Xavier had UEFI, Orin had better UEFI, and now this) -- which would imply it has full UEFI + ACPI like the Spark. But when I looked at the kernel in their Thor L4T release, it looked like it was still loaded with Jetson-specific SOC drivers on top of a heavy fork of the PREEMPT_RT patch series for Linux 6.8; I did not look too hard, but it still didn't seem ideal. Maybe you can probably boot a "normal" OS missing most of the actual Jetson-specific peripherals, I guess.
It's a bit ambiguous but I can't edit now, sorry. What I meant to say was that it boots using the same mechanism as x86 machines that you are familiar with, not that it is an x86 machine itself.
I assume the driver code just isn't in mainline linux and installing the correct toolchain isn't always easy. Having it turnkey available is nice and fundamentally new hardware just isn't going to have day 1 linux support.
You're free to lift the kernel and any drivers/libraries and run them on your distribution of choice, it'll just be hacky.
I ordered one that arrived last week. It seems like a great idea with horrible execution. The UI shows strange glitchy/artifacts occasionally as if there's a hardware failure.
Regarding limited memory bandwidth: my impression is that this is part of the onramp for the DGX Cloud. Heavy lifting/production workloads will still need to be run in the cloud.
CUDA is only on nvidia GPUs, I guess a RTX Pro 6000 would get you close, two of them are 192GB in total. Vastly increased memory bandwidth too. Maybe two/four of the older A100/A6000 could do the trick too.
Yes? Apple does not document their GPUs or provide any avenue for low-level API design. They cut ties with Khronos, refuse to implement open GPU standards and deliberately funnel developers into a proprietary and non-portable raster API.
Nvidia cooperates with Khronos, implements open-source and proprietary APIs simultaneously, documents their GPU hardware, and directly supports community reverse-engineering projects like nouveau and NOVA with their salaried engineers.
Pretty much the only proprietary part is CUDA, and Nvidia emphatically supports the CUDA alternatives. Apple doesn't even let you run them.
The resale cost shouldn't be ignored either, that Mac Studio will definitely resell for more than this will by a significant amount. Least of all because the Mac Studio is useful in all kinds of industries whereas this is quite niche.
Can you be a bit more specific what technology you're actually referring to? "Unified memory" is just a marketing term, you could mean unified address space, dual-use memory controllers, SOC integration or Northbridge coprocessors. All are technologies that Nvidia has shipped in consumer products at one point or another, though (Nintendo Switch, Tegra Infotainment, 200X MacBook to name a few).
They're basically describing the Jetson and Tegra lineup, then. Those were featured in several high-end consumer devices, like smart-cars and the Nintendo Switch.
For how shit it all is, it's still the easiest to use, with most available resources when you inevitable need to dig through stuff. Just things like nsight GUI and available debugging options ends up bringing together a better developer experience compared to other ecosystems. I do hope the competitors get better though because the current de facto monopoly helps no-one.
My reasons for not choosing an Apple product for such a use-case:
1- I vote with my wallet, do I want to pay a company to be my digital overlord, doing everything they can to keep me inside their ecosystem? I put too much effort to earn my freedom to give it up that easily.
2- Software: Almost certainly, I would want to run linux on this. Do I want to have something that has or eventually will have great mainstream linux support, or something with closed specs that people in Asahi try to support with incredible skills and effort? I prefer the system with openly available specs.
I've extensively used mac, iphone, ipad over time. The only apple device I ever bought was an ipad, and I would never buy it, if I knew they deliberately disable multitasking on it.
> container is a tool that you can use to create and run Linux containers as lightweight virtual machines on your Mac. It's written in Swift, and optimized for Apple silicon.
That would have been an impressive piece of technology in 2015, when WSL was theoretical. To release it in 2025 is a very bad trend, and it reflects Apple's isolation from competition and reluctance to officially support basic dev features.
Container does nothing to progress the state of supporting Linux on Apple Silicon. It does not replace macOS, iBoot or the other proprietary, undocumented or opaque software blobs on the system. All it does is keep people using macOS and purchasing Apple products and viewing Apple advertisements.
My hope was to find a system which does ASR, then LLM processing with MCP use and finally TTS: "Put X on my todo list" / "Mark X as done" -> LLM thinks, reads the todo list, edits the todo list, and tells me "I added X to your todo list", ... "Turn all the lights off" -> llm thinks and uses MCP to turn off the lights -> "Lights have been turned off". "Send me an email at 8pm reminding me to do" .... "Email has been scheduled for 8pm"
That's all I want. It does not have to be fast, but it must be capable of doing all of that.
Oh, and it should be energy efficient. Very important for a 24/7 machine.
You can already do that on most desktop GPU's (even going as far as prev gen Nv 1050/1060/1070 for example).
You'll need a model able to work with tools, like llama 3.2 (https://huggingface.co/meta-llama), serve it, hook up MCPs, include a STT interface, and you're cooking.
Even a bottom of the barrel N95 has audio acceleration features helping with speech to text, but the LLM inference part still will be far from being efficient.
Plus, you need to keep the card at "ready" state, you can't idle/standby it completely.
Why is every computer listing nowadays look the same with the glowing golden and blue chip images and the dynamic images that appear when you scroll down.
Please give me a good old html table with specs will ya?
These are primarily useful for developing CUDA targeted code on something that sits on your desk and has a lot of RAM.
They're not the best choice for anyone who wants to run LLMs as fast and cheap as possible at home. Think of it like a developer tool.
These boxes are confusing the internet because they've let the marketing teams run wild (or at least the marketing LLMs run wild) trying to make them out to be something everyone should want.
It's very, very good as an ARM Linux development machine; the Cortex-X925s are Zen5 class (with per-core L2 caches twice as big, even!) and it has a lot of them; the small cores aren't slouches either (around Apple M1 levels of perf IIRC?) GB10 might legitimately be the best high-performance Linux-compatible ARM workstation you can buy right now, and as a bonus it comes with a decent GPU.
A GPU cluster would work better but if you're only testing things out using CUDA and want 200GB networking and somewhat low power all in one this would be the device for you
AI stuff aside I'm frankly happy to see workstation-class AArch64 hardware available to regular consumers.
Last few jobs I've had were for binaries compiled to target ARM AArch64 SBC devices, and cross compiling was sometimes annoying, and you couldn't truly eat your own dogfood on workstations as there's subtle things around atomics and memory consistency guarantees that differ between ISAs.
Mac M series machines are an option except that then you're not running Linux, except in VM, and then that's awkward too. Or Asahi which comes with its own constraints.
Having a beefy ARM machine at my desk natively running Linux would have pleased me greatly. Especially if my employer was paying for it.
Even cheaper, unless you want the really high-end enterprise stuff. You can run ComfyUI pretty comfy for $0.30 to $0.40 per hour, if AI art is your goal.
Is there something similar with twice the memory/bandwidth? That's a use case that I would seriously consider to run any frontier open source model locally, at usable speed. 128GB is almost enough.
Fill up the memory with a large model, and most of your memory bandwidth will be waiting on compute shaders. Seems like a waste of $5,000 but you do you.
Looks like a pretty useful offering, 128Gb Memory Unified, with the ability to be chained. IN the Uk release price looks to be £2999.99 Nice to see AI Inference becoming available to us all, rather than using a GPU ..3090etc.
You'd have to be doing something where the unified memory is specifically necessary, and it's okay that it's slow. If all you want is to run large LLMs slowly, you can do that with split CPU/GPU inference using a normal desktop and a 3090, with the added benefit that a smaller model that fits in the 3090 is going to be blazing fast compared to the same model on the spark.
Eh, this is way overblown IMO. The product page claims this is for training, and as long as you crank your batch size high enough you will not run into memory bandwidth constraints.
I've finetuned diffusion models streaming from an SSD without noticeable speed penalty at high enough batchsize.
Asus make some really useful things, but the v1 Tinker Board was really a bit problem-ridden, for example. This is similarly way out on the edge of their expertise; I'm not sure I'd buy an out-there Asus v1 product this expensive.
This bit of the FAQ was such a non-answer to their own FAQ, you really have to wonder:
>> What is the memory bandwidth supported by Ascent GX10?
> AI applications often require a bigger memory. With the NVIDIA Blackwell GPU that supports 128GB of unified memory, ASUS Ascent GX10 is an AI supercomputer that enables faster training, better real-time inference, and support larger models like LLMs.
> I guess that's the kindest possible interpretation. The other interpretation is that the answer is not a good one.
If they wanted to do that, they should have just omitted the question from their FAQ. An evasive answer in a FAQ is a giant footgun, because it just calls attention to the evasion.
Funny to wakeup and see this on the front page - I literally just bought a pair last night for work (and play) somewhat on a whim, after comparing the available models. This one was available the soonest & cheapest, CDW is giving 100 off even, so 2900 pre tax.
Really interested to see if anyone starts using the fancy high end Connect-X 7 NIC in these DGX Spark / GB10 derived systems. 200Gbit RDMA is available & would be incredible to see in use here.
That would depend on your idea of "good". It would be an upstream swim in most regards, but you could certainly make it work. The Asahi team has shown that you can get steam working pretty well on ARM based machines.
But if gaming is what you're actually interested in, then it's a pretty terrible buy. You can get a much cheaper x86-based system with a discrete GPU that runs circles around this.
1) This still has raster hardware, even ray tracing cores. It's not technically an "AI focused card" like the AMD Instinct hardware or Nvidia's P40-style cards.
2) It kinda does have a stack. ARM is the hardest part to work around, but Box86 will get the older DirectX titles working. The GPU is Vulkan compliant too, so it should be able to leverage Proton/DXVK to accommodate the modern titles that don't break on ARM.
The tough part is the price. I don't think ARM gaming boxes will draw many people in with worse performance at a higher price.
Memory bandwidth is a joke. You would think by now somebody would come out with a well balanced machine for inference instead of always handicapping one of the important aspects. Feels like a conspiracy.
At least the m5 ultra should finally balance things given the significant improvements to prompt processing in the m5 from what we've seen. Apple has had significantly higher memory bandwidth since the m1 series approaching 5 years old now. Surely an nvidia machine like this could have at bare minimum 500Gb+ if they cared in the slightest about competition.
is this another product they're pushing out for publicity. I mean how much testing has been done for this product. Need more specs and testing results to illuminate capabilities, practicality.
If you touch the image when scrolling on mobile then it opens when you lift your finger. Then when you press the cross in the corner to close the image, the search button behind it is activated.
How can a serious company not notice these glaring issues in their websites?
Taiwanese companies still don't value good software engineering, so talented developers who know how to make money leave. This leaves enterprise darlings like Asus stuck with hiring lower tier talent for numbers that look good to accounting.
> What is the memory bandwidth supported by Ascent GX10? AI applications often require a bigger memory. With the NVIDIA Blackwell GPU that supports 128GB of unified memory, ASUS Ascent GX10 is an AI supercomputer that enables faster training, better real-time inference, and support larger models like LLMs.