The Asus Ascent GX10 is a MiniPC with supercomputer ambitions for AI developers – but it's not cheap

Unless you've been living under a rock for the past 18 months, it's impossible not to be aware of how many of its eggs Nvidia has put in the AI basket.

To further stoke the AI fire, Nvidia released a dedicated desktop box aimed at AI developers called the DGX Spark, which combines Nvidia's latest generation of ARM CPU, GPU, and networking technologies in one user-friendly and affordable package.

The phrases "user-friendly" and "affordable" are doing a lot of lifting in that previous paragraph, but given the lack of alternatives if you want to develop your AI project locally, this is as good as it gets.

Image 1 of 9

Nvidia has also partnered with several OEMs to bring variations on the Spark theme to market. All the OEM boxes share the same basic specification as the Nvidia Ur-Box, suggesting Nvidia is keeping everyone on a pretty tight leash.

Dell has released the Dell Pro Max with GB10, Gigabyte the AI Top Atom. Lenovo with the ThinkStation PGX, MSI with its EdgeXpert MS-C931, and finally Asus with the Ascent GX10.

It's the Asus Ascent GX10 we are looking at today. All the variations on the DGX Spark will set you back around £4,000, depending on the amount of storage it comes with; the Ascent we are poking with a stick can be picked up in the UK for £3,640.

Asus Ascent GX10: Design

Given the esoteric hardware that lurks inside the Ascent GX10 is impressively small, stylish, and unobtrusive. It also has one major advantage over Nvidia's own box and the other OEM competition in that the power switch is at the front rather than the back or, in Dell's case, the top.

That may not be a big deal if you just have one box sat next to your monitor, but if you have several arranged in a rack, and that is very much how this system is intended to be used, then you'll be left groping around the back to power your system up.

Physically, the Ascent GX10 is a 150 x 150 x 51mm metal box weighing 1.48Kg. The front, back, and side panels are painted in matte slider while the top has a cross-hatched mirror finish.

The wavey slats on the front presumably provide more surface area for cooling, but as essentially an aesthetic consideration. Compared to the rather baroque DGX Spark and the utilitarian offerings from the other OEMs, the Asus box is by far the most visually appealing.

To be fair to Nvidia, the external design of the DGX Spark is very much an homage to the 2016 DGX1, a $130,000 61Kg lump the size of a large suitcase with a paltry-by-modern-standards 170TFLOPS capability.

All the ports are grouped at the rear below the large air vent. Running left to right, you'll find a Kensington lock, four 20Gbps (Gen 2×2) USB-C ports, one of which serves as the connection for the 240W power brick, an HDMI 2.1 video output, 10GbE Ethernet, and two 200Gbps ConnectX-7 QSFP112 ports.

Incidentally, the three USB-C ports that are not powering the box also support DisplayPort 2.1 DP alt-mode video output. Another consideration is that the GB10 chip has a TDP of 140W, which leaves 100W for all the other bits and bobs, including the ConnectX-7 NIC and the three USB ports. Wireless communications are handled by a MediaTek card that supports Wi-Fi 7 and Bluetooth 5.4.

Internal access is minimal, but given that everything important is soldered in place, this is not the end of the world. Unscrew the base grille and then another metal plate, and you can easily get at the 2242 SSD, which means you can buy the cheapest version of the box and upgrade the storage should the need arise.

Dig deep into the Asus product sheets, and you'll find that only the 4TB configuration ships with a PCIe 5.0 x 5 SSD. The smaller SSDs are PCIe 4.0 x 4 specification.

Like the Nvidia DGX Spark and all the other OEM iterations thereof, the Ascent GX10 runs on NVIDIA's DGX OS, a variant of Ubuntu 24.04 LTS (for ARM) Linux.

The primary differences between DGX OS and vanilla Ubuntu are really just a matter of the apps Nvidia has added, such as the dashboard to manage your AI projects, Nvidia Sync utility that manages back-end SSH keys and tunnelling, and all the relevant drivers.

The OS dashboard also directly links to Nvidia's DGX Spark Playbooks repository on GitHub. These provide a great selection of programmes for getting a wide variety of AI models up and running, including ComfyUI, Pytorch, and vLLM, to name but three.

These playbooks are an important part of the GX10's appeal to those who are rather new to the world of AI development, and one reason why the GX10 or any of its Spark cousins are perhaps a better bet for AI developers than the likes of a Mac Studio or a do-it-yourself rig built around an Nvidia discrete GPU or two.

Asus Ascent GX10: Specs and Performance

Running the show inside the GX10 is an Nvidia GB10 ARM v9.2-A CPU, which consists of 10 Arm Cortex-X925 and 10 Arm Cortex-A725 cores, a Blackwell GB20B GPU. You also get 128GB of soldered LPDDR5X-8533 RAM, a brace of ConnectX-7 200Gbps networking ports, and a 4GB 2242-format SSD.

Versions with a 1GB or 2GB SSD are also available for rather less money, but the majority of the boxes on sale through retail channels are the 4GB variety, and given the intended use, that would surely be the model to go for.

Before we press on, it's worth making clear that one reason these Nvidia-system boxes are so expensive is because of that Nvidia ConnectX-7 interface, which easily accounts for upwards of one-third of the asking price.

It goes without saying that 200Gbps of bandwidth is nothing like what you get from a full-fledged GB200/GB300 server, but it does provide developers with access to more power and also a way to see how their models perform on a scaled setup.

Measuring performance is borderline superfluous on a machine that boasts 1 petaFLOP (that's 1,000 TFLOPS in old money) of performance, and our testing regime isn't suited to measuring the exact potential of GX10 to run large LLMs and VLMs at the same time or create local AI agents.

Asus's demonstrations undoubtedly confirmed Nvidia's performance claims, such as the impressive abilityfine-tuningning Pytorch LoRA llama3 8B at 53K peak tokens per second (TPS) and generating a 1K FP16 image in Flux1.dev 12b in double quick time thanks to having ready access to the 35GB of RAM that the job required.

In a rather unscientific test, we ran some basic prompts in LM Studio on both the GX10 and the Framework Studio. The Asus box averaged 4.67 tokens per second while the Framework came in at 13.4 TPS. Of course, our Framework box only has 32GB of RAM.

Running some 1K image creation tasks in ComfyUI gave the Asus box an even larger advantage, the GX10 taking 216 seconds to do a job that the Framework box took over 400 seconds.

At a more general level, the GB10 CPU scored 3,104 in the GeekBench 6 single-core benchmark and 20,048 in the multi-core test. For comparison, a 2024 Mac Mini running on the M4 Pro CPU scored 3,986 and 22,857, respectively, while the AMD Ryzen AI Max 385 CPU in the Framework Desktop scored 2,952 and 16,599.

Throughout all the test the GX10 impressed by how quiet it was, the dual fan system seldom making much more than a loud whisper. Under heavy stress, the external case got a little warm in places, but by no more than your average laptop when under duress.

Our takeaway is that if you are developing large AI frameworks and want to do so locally, then the GX10 and its relatives are the smallest, fastest, and most convenient way to do it.

Granted, you could build yourself a development rig based on a couple of Blackwell gaming GPUs and potentially get more performance, but the result would be bigger, louder, far more power hungry, and probably more expensive when you factor in the cost of RAM, a CPU, a box, and everything else, let alone a comparable network card.

Asus Ascent GX10: Is it worth it?

With an item as esoteric as the Ascent GX10, that's not an easy question to answer. For a substantially smaller pile of cash, you can buy an Apple Mac Studio or a system built around one of AMD's Ryzen AI Max chips, like the Framework Desktop modular PC we looked at recently or HP's new Z2 Mini G1a Workstation, both of which can be configured with 128GB of RAM.

Of course, none of those machines have two 200GbE OSFP112 networking ports, and that's the first area in which the GX10 excels: scalability. Buy two of these things and hook them together, and you will really be cooking with gas. Though you'll also be over £7,000 out of pocket.

The value of the various Nvidia AI playbooks shouldn't be overlooked either. Having some GB10-specific workbooks is something we think will prove useful to developers wanting to further hone and expand their AI capabilities.

The Ascent GX10, then, is very much a thoroughbred designed for a specific job. Many developers will be happy with one of the forthcoming AMD Ryzen Max+ boxes, which are not only cheaper but run Windows 11, making them more versatile, but if you are deadly serious about your AI development and actually need the performance that more than one GX10 box can deliver, then it's a pretty impressive device for the asking price.

Asus Ascent GX10 Specification