No, the traditional AI desktop isn't dead. The ASUS ProArt GR1X fits 64 GB or 128 GB of unified memory into a box of about 1.1 litres, with a 140 W sustained power rating. That lets a desk-sized machine hold models that used to need a large graphics card or a tower. But memory capacity is only one part of AI performance. A conventional tower with a desktop RTX GPU still has several times the memory bandwidth, much more raw GPU power, real expansion slots and a mature software stack. The likely outcome is that both kinds of machine find their own jobs, not that one replaces the other.
How does a compact RTX Spark desktop compare with a GPU tower?
| Factor | ASUS ProArt GR1X (RTX Spark N1X) | Traditional tower (desktop RTX GPU) |
|---|---|---|
| Memory for AI | 64 GB or 128 GB LPDDR5X, shared by CPU and GPU | 32 GB (RTX 5090) or 96 GB (RTX PRO 6000) per card, plus system RAM |
| Memory bandwidth | About 300 GB/s (reviewers' calculation; NVIDIA hasn't published a figure) | 1,792 GB/s on an RTX 5090 or RTX PRO 6000 Workstation Edition |
| GPU class | Between an RTX 5060 Laptop and RTX 5070 Laptop GPU in 3DMark (Notebookcheck, laptop N1X) | Up to the fastest desktop cards NVIDIA makes |
| Power | 140 W sustained TDP, 240 W USB-C adapter (ASUS) | 575–600 W for the GPU alone, plus CPU and platform |
| Expansion | Two M.2 SSD slots; memory soldered; no PCIe card slot | PCIe slots, upgradeable RAM, more storage, multiple GPUs |
| Software | Windows 11 on Arm; CUDA on Windows on Arm is a developer preview | Windows or Linux on x86 with the full CUDA stack |
| Size | 150 × 150 × 51 mm, about 1.1 L, 1.48 kg | Mid or full tower |
Source: ASUS ProArt GR1X product and techspec pages (accessed 8 Nov 2026); NVIDIA RTX Spark specifications; Notebookcheck N1X GPU page (15–16 Oct 2026); NVIDIA RTX 5090 and RTX PRO 6000 specifications. No independent GR1X benchmark had been published at the time of writing, so GR1X performance is inferred from N1X laptops.
What does the GR1X actually change?
It brings a lot of memory onto a desk at low power. Until recently, running a model the size of gpt-oss-120b (about 65 GB of weights) locally meant an RTX PRO 6000-class card or splitting the model across several GPUs. An RTX Spark machine with 128 GB can hold it in one memory pool. Techtesters measured about 63 tokens per second on gpt-oss-120b on an MSI Prestige N16 Flip AI+ laptop with the same chip, sustained over ten minutes (16 Oct 2026, quantisation and tool not published).
PCWorld's ProArt P16 review shows where this matters. A gpt-oss-120b research workload took 33 minutes 13 seconds on the 128 GB laptop and 52 minutes 51 seconds on an RTX 5090 desktop, because the 5090 ran out of its 32 GB of VRAM and had to lean on system memory (29 Oct 2026). PCWorld also noted that the 5090 is the faster machine in general.
The second change is the "agent computer" idea. NVIDIA says RTX Spark desktops are "built to run personal AI agents 24/7 right at your desk", and ASUS says the GR1X is "engineered for 24/7 agentic AI". A small, quiet box that stays on and runs a local model for agents is a real new category. Our post on whether you need a dedicated AI agent computer looks at when that's worth paying for.
Where does the traditional desktop still win?
On anything that fits in a desktop GPU's memory. Generating text is limited mainly by memory bandwidth, because the model's weights are read for every token. A desktop RTX 5090 or RTX PRO 6000 reads memory about six times faster than the N1X. For a dense 32B model at 4-bit (about 20 GB), the theoretical ceiling at 300 GB/s is roughly 15 tokens per second. The same model on a 1,792 GB/s card has a ceiling several times higher (our arithmetic; real speeds are lower on both).
Published laptop results point the same way. Tom's Guide found the Surface Laptop Ultra's Procyon LLM score below a desktop RTX 5060 Ti 16 GB (Llama 3.1 8B: 2,173 vs 2,642, 15 Oct 2026). Notebookcheck rated the N1X GPU between an RTX 5060 Laptop and an RTX 5070 Laptop GPU in 3DMark. The GR1X's 140 W budget is higher than a laptop's 45–80 W, which may help compute-heavy work, but no desktop result had been published at the time of writing.
Towers also remain the choice for training and fine-tuning large models, GPU rendering at scale, multi-GPU work and high-end gaming. Notebookcheck reported that Call of Duty, F1 25 and Battlefield 6 would not run on N1X systems in October, and Microsoft has committed Call of Duty support for 2027.
The GR1X makes a big model fit on your desk. It doesn't make a big model fast, and it doesn't make a small model faster than a desktop GPU would.
Why does expandability still matter?
Because AI requirements keep moving. On the GR1X, the LPDDR5X memory is soldered, so the 64 GB or 128 GB you buy is what you keep. You can replace or add SSDs in its two M.2 slots, but there's no PCIe slot for a second GPU and no ConnectX port for fast clustering. A tower can take more RAM, a bigger GPU next year, a second card, or a faster network adapter. When a part fails, it can usually be swapped individually.
That flexibility costs money, power and desk space. If your workload will stay within 128 GB, a fixed design may be fine; if models or users will grow, the tower's upgrade path is worth a lot. See compact unified memory vs expandable GPU towers.
What about software?
This is the biggest reason towers aren't going away yet. As of November 2026:
- CUDA on Windows on Arm is a developer preview, and NVIDIA warns that some PyTorch workflows can trigger GPU timeouts.
- vLLM has no native Windows support (WSL2 or an unofficial port only), and TensorRT-LLM's install documentation is Linux-only.
- Windows gives the GPU a fixed share of memory. On 64 GB N1X systems, XDA found at most about 55 GB exposed to the GPU.
- NVIDIA has not committed to Linux drivers for the N1X.
A Linux tower, or a Linux GB10 box such as the ASUS Ascent GX10, runs the standard data-centre toolchain today. Our Windows on Arm app compatibility guide covers the application side.
Is the market moving away from towers?
There's no evidence of that yet. RTX Spark desktops were only arriving in late 2026, and none had an independent benchmark published by early November. Jon Peddie Research reported the workstation market up 12% year on year in Q1 2026 and down 2.3% in Q2, and its analyst urged caution about reading too much into one weak quarter. Microsoft, Dell, HP, MSI and ASUS have all announced compact RTX Spark machines, but alongside their towers, not instead of them.
A fairer reading is that compact unified-memory boxes are taking over some jobs towers did expensively, such as running one large model quietly on a desk. Jobs that need bandwidth, scale or Linux software stay on towers and servers.
So who should buy which?
- A GR1X-class box suits one person or a small team running large models for inference, always-on agents, or creative work with the bundled ASUS tools, on Windows, where low power and noise matter.
- A GPU tower suits anyone who needs the fastest token generation on models that fit in VRAM, fine-tuning or training, GPU rendering, multi-GPU growth, or Linux-first tools.
- Both is common. Many teams will put a compact box beside a tower or server rather than choose one. For model sizing, see the memory buying mistake; for workloads that need a dedicated GPU workstation, see why the GR1X might be the wrong choice for serious AI developers.
Buying in Singapore
The ASUS ProArt GR1X is listed as price on request. ASUS hasn't published a price; as a SourceIT estimate based on comparable RTX Spark devices, not an official ASUS price, expect roughly S$6,200–7,200 for 64 GB / 1 TB and S$9,000–12,000 for the 128 GB models, including GST. For a tower route, the RTX PRO 6000 Blackwell Workstation Edition is S$29,100 (listed 8 Nov 2026) before the host system; our RTX PRO 6000 workstation build guide covers the rest. For a compact Linux alternative, the ASUS Ascent GX10 128 GB / 1 TB is S$10,350 (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
Can the ASUS ProArt GR1X replace a desktop workstation? For running large models for inference and for always-on agents, it can replace a tower for some users. For fine-tuning, GPU rendering, multi-GPU work or the fastest token generation on models that fit in a desktop GPU, a tower is still the better tool.
Is the GR1X faster than a desktop RTX GPU? Only when the model doesn't fit in the desktop card's VRAM. When it does fit, a desktop RTX 5090 or RTX PRO 6000 has about six times the memory bandwidth and generates text faster.
Can I upgrade the GR1X later? You can change or add SSDs in its two M.2 slots. The memory is soldered and there's no PCIe slot for another GPU.
Does the GR1X run Linux? It ships with Windows 11 on Arm. NVIDIA has not committed to Linux drivers for the N1X chip; if you need Linux and CUDA in a compact box, look at a GB10 system such as the ASUS Ascent GX10.
Are there GR1X benchmarks yet? Not at the time of writing. Performance figures here come from laptops with the same N1X chip, which run at lower power than the GR1X's 140 W.
