It depends entirely on whether your model fits in 32 GB. A desktop GeForce RTX 5090 has 32 GB of GDDR7 at 1,792 GB/s, about six times the memory bandwidth of the ASUS ProArt GR1X's RTX Spark N1X chip (ASUS quotes up to 300 GB/s). Anything that fits on the 5090 runs several times faster there. Once a model outgrows 32 GB, the 5090 has to spill into much slower system RAM, and the GR1X's 128 GB of unified memory starts to win. PCWorld saw exactly that with gpt-oss-120b. So more memory beats more GPU power only for the models that need the memory.
How do they compare on paper and in published tests?
| Item | ASUS ProArt GR1X (N1X) | Desktop PC with RTX 5090 |
|---|---|---|
| Memory for AI | 64 or 128 GB unified, shared with the CPU | 32 GB GDDR7 on the card, plus system RAM |
| Memory bandwidth | Up to 300 GB/s (ASUS) | 1,792 GB/s |
| Power | 140 W sustained platform, 240 W adapter | 575 W for the card alone; NVIDIA requires a 1,000 W system |
| Qwen3 32B 4-bit, tokens/s | 11.8–12.1 (Lenovo Yoga Pro 9n 64 GB, same chip, Ollama) | 43 (llama.cpp) |
| gpt-oss-20b, tokens/s | 58–59 (Surface Laptop Ultra, same chip, Ollama) | 282.5 (llama.cpp) |
| gpt-oss-120b agentic workload, same prompts | 33 min 13 s (ASUS ProArt P16 128 GB, same chip) | 52 min 51 s (RTX 5090, i7-14700K, 96 GB RAM) |
| Size | 1.1 litres, 1.48 kg | Full tower |
Source: ASUS GR1X product page (8 Nov 2026); NVIDIA RTX 5090 specifications; XDA (26 Oct 2026); StorageReview Surface Laptop Ultra review (Oct 2026); hardware-corner (Mar 2026) and the llama.cpp gpt-oss guide for RTX 5090 figures; PCWorld ProArt P16 review (29 Oct 2026). No GR1X benchmark had been published at the time of writing, so the GR1X column uses RTX Spark laptops with the same chip at lower power. Tools and settings differ between sources, so compare the rough size of the gap, not exact numbers.
When does the RTX 5090 win?
Whenever the model, its context and any other working data fit in 32 GB. Token generation is limited by how fast the GPU can read the model's weights, and the 5090 reads them about six times faster. The rows above show the pattern: roughly 3.5x on a 32B model and about 5x on gpt-oss-20b. The 5090 also has more raw compute for prompt processing, image generation and video generation, and the CUDA software stack on x86 Windows and Linux is mature in a way Windows on Arm isn't yet.
Typical 5090 territory:
- 8B–32B chat and coding models at 4-bit, with moderate context
- Image generation with FLUX-class and SDXL models
- Serving a small model to several people at once, where its bandwidth and compute keep throughput high
- Gaming, rendering and anything else that needs a full desktop GPU
When does the GR1X's memory win?
When the model doesn't fit in 32 GB. gpt-oss-120b's weights are about 65 GB, so on a 5090 most of the model sits in system RAM and crosses the PCIe link for every token. Published offload results for gpt-oss-120b on a 5090 range from about 10 to 30 tokens per second depending on the system's RAM and settings. RTX Spark laptops with 128 GB keep the whole model in unified memory and measured around 59–63 tokens per second.
That's what PCWorld's test captured. Running the same gpt-oss-120b research workload, the ASUS ProArt P16 128 GB finished in 33 minutes 13 seconds against 52 minutes 51 seconds for an RTX 5090 desktop, because the 5090 ran out of video memory. PCWorld added that the 5090 is the faster GPU in general.
A fast GPU with too little memory and a big memory pool with modest bandwidth fail in opposite ways. Pick the one whose weakness your models won't hit.
What about long contexts and big image or video models?
Context length moves models across the 32 GB line. A 32B model at 4-bit takes about 20 GB of weights, which leaves room on a 5090 for a moderate context but not for 100K-plus tokens or several users at once. The GR1X has the headroom, though its default GPU memory split on Windows is about half of the installed memory, and on 64 GB units XDA found at most 55.4 GB exposed to the GPU. Video models such as Wan 2.2 14B are similar: at full precision they want far more than 32 GB, so 5090 users rely on 8-bit or GGUF versions and block swapping. Our RTX 5090 32 GB limits post covers the developer side of that ceiling.
What do power and price look like?
The GR1X's whole platform is rated at 140 W sustained. An RTX 5090 card alone has a 575 W rating, and reviewers measured about 463–546 W in games. At the Q4 2026 SP Group tariff of 31.16¢/kWh including GST, every extra 400 W for 8 hours a day on 250 working days adds about S$249 a year. That's our illustrative arithmetic, not a measured figure.
On price, the RTX 5090 card's US launch price is US$1,999, but late-2026 US street prices were roughly US$6,500–8,500, and SourceIT had no 5090 card in stock at the time of writing. The GR1X's price hadn't been announced. Our estimate, based on comparable RTX Spark devices and not an official ASUS price, is about S$9,000–10,800 for 128 GB/2 TB. That's a complete computer, against a card that still needs a high-end PC around it.
Which should you choose?
- Your models fit in about 24–28 GB with room for context: an RTX 5090 desktop (or the laptop-class alternative below) is much faster.
- You run 70B–120B models, especially mixture-of-experts like gpt-oss-120b: the GR1X's 128 GB is the better fit, at a fraction of the power. Read our memory trap post first so speeds don't surprise you.
- You need both speed and memory: neither. Look at the RTX PRO 6000's 96 GB at the same 1,792 GB/s, covered in our GR1X vs RTX PRO 6000 post.
The published GB10 vs RTX 5090 comparison makes the same argument for the DGX Spark, whose memory bandwidth is similar to the GR1X's.
Buying in Singapore
SourceIT lists the ProArt GR1X as price on request in the ASUS ProArt GR1X collection. RTX 5090 graphics cards weren't in stock at the time of writing, so ask us about availability. If you want 5090-class graphics in a compact system now, the ASUS ROG NUC 16 Edition 20 uses the RTX 5090 Laptop GPU (24 GB, a smaller chip than the desktop card) and is S$11,888.40 (listed 8 Nov 2026). Prices change, so ask for a current quotation. To get a quote, 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 RTX 5090 faster than the ASUS ProArt GR1X for AI? Yes, for models that fit in its 32 GB. Its memory bandwidth is about six times higher, and published tests show roughly 3.5–5x faster text generation on 20B–32B models compared with RTX Spark laptops on the same chip as the GR1X.
When is the GR1X faster than an RTX 5090? When the model is too big for 32 GB. PCWorld's gpt-oss-120b workload finished in 33 minutes on an RTX Spark ProArt P16 against 53 minutes on an RTX 5090 desktop, because the 5090 ran out of video memory.
Can an RTX 5090 run gpt-oss-120b at all? Yes, by offloading part of the model to system RAM, but published results range from about 10 to 30 tokens per second depending on the system. The model's roughly 65 GB of weights don't fit in 32 GB.
Which uses less power? The GR1X, by a wide margin. Its platform is rated at 140 W sustained, while an RTX 5090 card alone is rated at 575 W and needs a 1,000 W system.
What if I need both lots of memory and high speed? Consider an RTX PRO 6000 Blackwell, with 96 GB at 1,792 GB/s, or a Mac Studio M5 Ultra with up to 512 GB at 1.2 TB/s. Both cost considerably more than the GR1X.
