It's real hardware with stretched marketing. The ASUS ProArt GR1X puts NVIDIA's RTX Spark N1X chip and up to 128 GB of unified memory in a 1.1-litre box, and with 128 GB it can genuinely run models of around 120 billion parameters that a 32 GB graphics card can't load at all. But the headline numbers on the spec sheet (1 petaflop, 120B parameters, 1 million tokens of context, 100+ fps gaming) are all best-case figures with the conditions left out. Read them as upper limits, not as what you'll see at your desk.
Which GR1X claims need a closer look?
| Claim (ASUS or NVIDIA) | What it means | What's left out |
|---|---|---|
| "Up to 1 PFLOP (FP4)" | Peak 4-bit compute of the Blackwell GPU | No sparsity note on the RTX Spark page; NVIDIA's DGX Spark page footnotes the same figure as using sparsity. Text generation speed depends on memory bandwidth, not this figure |
| "Runs 120B-parameter LLMs locally, up to 1M tokens of context" | A 120B-class model fits in 128 GB | No quantisation, model or speed given. Doesn't apply to the 64 GB version |
| "Renders 3D scenes up to 90GB" | Large scenes can sit in unified memory | Needs the 128 GB model and most of the memory assigned to the GPU |
| "AAA Gaming 1440p 100+ FPS" | Some games reach 100 fps at 1440p | No game, settings or upscaling named |
| "1.6X more efficient cooling" | Better cooling than something | No baseline stated |
| "Built for 24/7" | Designed for always-on use | No duty-cycle term published in a warranty |
| "Literally runs everything the world has ever created" (NVIDIA CEO, Computex) | Broad Windows app support through emulation | Call of Duty, F1 25 and Battlefield 6 would not run on RTX Spark laptops in Notebookcheck's testing |
Source: ASUS ProArt GR1X product page and press releases (2 Sep and 8 Oct 2026); ASUS gaming slides via Wccftech (9 Oct 2026); NVIDIA RTX Spark and DGX Spark product pages; Notebookcheck (16–18 Oct 2026); Windows Central (1 Jun 2026). All claims are vendor statements, not independent test results.
Is the GR1X really an "AI supercomputer"?
The label gets used loosely across this class of machine. NVIDIA calls the DGX Spark a personal AI supercomputer, and ASUS sells its GB10-based Ascent GX10 the same way. The GR1X belongs to the same 1 petaflop FP4 class.
For scale, NVIDIA rates the DGX Station GB300, its deskside workstation, at 20 petaflops of FP4 with sparsity. That's twenty times the GR1X's headline figure, with 748 GB of memory instead of 128 GB. Traditional supercomputers are also measured in double-precision maths for simulation, which neither machine is built for.
A fairer description is "a compact AI PC with an unusually large pool of GPU-accessible memory". That's still a meaningful thing to own. It just isn't a data-centre system shrunk to fit on a desk.
What does 1 petaflop mean for real AI speed?
Very little for chat-style use. A petaflop rating describes how much maths the GPU can do per second at 4-bit precision, with sparsity. On Blackwell, the sparse rate is typically double the dense rate.
Generating each new token of a reply, though, means reading the model's active weights from memory, so the limit is memory bandwidth. ASUS lists the GR1X at up to 300 GB/s. As a rough ceiling, a dense 27B model at 4-bit (about 14 GB) can't exceed about 21 tokens per second at that bandwidth, and a 70B model at 4-bit (about 42.5 GB) can't exceed about 7. Real results land below those ceilings. Compute matters more for reading long prompts, image generation and rendering. We go through the maths in why 1 petaflop doesn't tell you how fast your AI will run.
Does "120B parameters and 1M tokens" hold up?
Partly. OpenAI's gpt-oss-120b is the typical 120B-class model. Its native 4-bit weights take about 65 GB, so it fits in a 128 GB GR1X but not in the 64 GB version. XDA found that a 64 GB RTX Spark laptop exposes at most about 55 GB to the GPU.
The 1 million token context is the stretch. For gpt-oss-120b, a 1-million-token conversation needs roughly 39 GB of extra memory for its cache at 16-bit precision, or about 19 GB at 8-bit. That brings the total to about 104 GB or 85 GB.
On RTX Spark laptops, Windows gives the GPU about half the memory by default. PCWorld found the ProArt P16 128 GB could be set to give up to 96 GB to the GPU through MyASUS; the GR1X's page doesn't document this. So 1 million tokens is possible only with an 8-bit cache and the memory split raised.
Even then, the model has to read the whole prompt first. At the 1,169–1,725 tokens per second that Ollama and NVIDIA measured for gpt-oss-120b prompt reading on a DGX Spark, a different system running Linux, 1 million tokens would take roughly 10–15 minutes before the first word of the answer.
Our earlier post on what NVIDIA's 120B and 200B model claims really mean covers the quantisation assumptions behind these figures.
The GR1X's spec sheet tells you what will fit. It doesn't tell you how fast it will run, and for most buyers that's the more important question.
What performance is realistic?
No GR1X benchmark had been published at the time of writing. The closest data comes from laptops with the same N1X chip, running at 45–80 W rather than the GR1X's 140 W. Techtesters (16 Oct 2026, quantisation not published) measured:
- gpt-oss-120b: 63.3 tokens per second on the MSI Prestige N16 and 58.9 on the ASUS ProArt P14.
- Qwen3.6-35B-A3B: 78.1 and 74.5.
- Qwen3.5-122B-A10B: 30.6 and 31.5.
- Qwen3.8-27B, a dense model: 15.8 and 15.
The pattern is that models which activate only a few billion parameters per token (mixture-of-experts models) run well, and dense models run much slower. The GR1X's extra power headroom may help prompt reading and image work. It is unlikely to change token generation much, because every N1X system has the same memory bandwidth. That's our reading of the specifications, not a measured result.
Gaming is the same story. HardwareZone measured about 75 fps in Cyberpunk 2077 at 1440p medium on the ProArt P16, and Notebookcheck found several popular games that wouldn't run. "100+ fps" will depend on the game, settings and upscaling.
So is it clever marketing or a real AI machine?
Both:
- Real: 128 GB of memory the GPU can use, a Blackwell GPU with FP4 support, in a small desktop that ASUS describes as near-silent. For running 120B-class mixture-of-experts models privately, that combination is rare.
- Marketing: "supercomputer", the petaflop headline and the 1-million-token context, which all describe limits rather than everyday speed. The software that makes NVIDIA strong is still in preview on Windows on Arm; see the seven GR1X limitations to know first.
If the claims you care about are memory-related, the 128 GB GR1X delivers. If they're speed-related, wait for independent desktop benchmarks or compare carefully with higher-bandwidth options. The 128 GB memory trap explains why the two aren't the same thing.
Buying in Singapore
The ASUS ProArt GR1X is in our ProArt GR1X collection at price on request. ASUS hasn't announced pricing; SourceIT estimates about S$6,200–7,200 for 64 GB and S$9,000–10,800 for 128 GB with 2 TB, based on comparable RTX Spark devices, not an official ASUS price. If you want to see the N1X chip in action now, the ASUS ProArt P16 128 GB laptop is S$10,773.20, and the Linux-based NVIDIA DGX Spark 4 TB is S$9,300 (both listed 8 Nov 2026). Prices change, so ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Is the ASUS ProArt GR1X a supercomputer? Not in the traditional sense. It's a compact AI PC rated at up to 1 petaflop of 4-bit compute, the same class as the DGX Spark. NVIDIA's DGX Station GB300 is rated at 20 times that.
Can the GR1X really run a 120B model? The 128 GB version can hold gpt-oss-120b, whose 4-bit weights are about 65 GB. The 64 GB version can't, because Windows exposes only about 55 GB to the GPU on 64 GB RTX Spark machines.
Is 1 million tokens of context usable? Only with an 8-bit cache and more memory assigned to the GPU, and reading a 1-million-token prompt could take 10–15 minutes based on DGX Spark measurements.
How fast is the GR1X at generating text? No GR1X benchmark has been published. Laptops with the same chip ran gpt-oss-120b at about 59–63 tokens per second in Techtesters' tests, and dense 27B models at about 15.
Will the GR1X run all my Windows apps? Most x64 apps run through emulation, but drivers and anti-cheat software need native Arm versions, and some games don't run. Check your key apps first.
