ASUS ProArt GR1X: Why 1 Petaflop Doesn't Tell You How Fast Your AI Will Run

Because "1 petaflop" measures how many low-precision calculations the chip can do at peak, while the speed you feel in a chatbot is mostly set by how fast it can read the model from memory. ASUS quotes "up to 1 PFLOP (FP4)" for the ProArt GR1X, the same headline NVIDIA uses for its DGX Spark, where the figure is footnoted as a theoretical peak using sparsity. When a local model writes an answer one token at a time, the GR1X's roughly 300 GB/s of memory bandwidth is usually the limit, not its petaflop. Compute matters more for reading long prompts, generating images and batching many users.

How do petaflops and memory bandwidth compare across machines?

Source: NVIDIA RTX Spark and DGX Spark product pages (read 8 Nov 2026); NVIDIA developer blog (24 Oct 2025); StorageReview RTX PRO 6000 review (7 Oct 2025); Techtesters MSI Prestige N16 Flip AI+ test (16 Oct 2026, quantisation not published); Apple Mac Studio specifications; oMLX community benchmark (24 Sep 2026, unverified). No GR1X benchmark had been published at the time of writing. Software and settings differ by row, so treat this as indicative.

The DGX Spark and the RTX PRO 6000 make the point. The PRO 6000 has four times the FP4 compute rating and about six and a half times the bandwidth, and it generates about three times as fast. The Mac Studio has no comparable petaflop rating at all, yet its 1.2 TB/s of bandwidth puts it ahead of both NVIDIA desktop chips on this test.

What does "1 petaflop FP4" actually mean?

A petaflop is a thousand trillion operations per second. FP4 is a 4-bit number format, the lowest precision in common use, mainly for running quantised models. Peak FP4 figures are the biggest numbers a chip can quote.

There's another detail. NVIDIA's DGX Spark page footnotes the same 1 petaflop as "theoretical FP4 TOPS using the sparsity feature". Sparsity is a technique that skips some zero values and can double the counted rate, but only for models prepared for it. NVIDIA's RTX Spark page omits that footnote, so we can't confirm the basis for the N1X's figure, but it's the same class of number. Dense FP4 throughput is commonly about half the sparse figure, and higher-precision formats such as FP8 and BF16, used for training, are lower again. We audit the rest of the marketing in real AI supercomputer or clever marketing.

What actually sets tokens per second?

Generating each new token means reading the model's active weights from memory. A useful rough ceiling for one user is:

tokens per second ≈ memory bandwidth ÷ bytes read per token

• Dense 70B at 4-bit (about 42.5GB of weights): 300 ÷ 42.5 ≈ 7 tokens per second at best.
• Qwen3 32B at 4-bit (about 19.8GB): about 15 tokens per second at best. XDA measured 11.8–12.1 on a same-chip Lenovo Yoga laptop, close to 80% of that ceiling.
• gpt-oss-120b is a mixture-of-experts model with only about 5.1B of its 117B parameters active per token. By our rough estimate it reads around 3GB per token, giving a ceiling near 100 tokens per second. The measured 63 tokens per second fits that.

These are our calculations from published model sizes and the reviewer bandwidth figure. Real results come in below the ceiling because of the conversation memory (the KV cache), software overhead and how well the runtime uses the hardware.

For local chat, look at memory bandwidth and active parameters first. Petaflops tell you about prompt reading and image generation; bandwidth tells you how fast the words appear.

So when does the petaflop matter?

• Reading long prompts (prefill): processing a long document before answering is compute-heavy. NVIDIA measured DGX Spark prefill on gpt-oss-120b at 1,725 tokens per second, against 55.4 for generation. HardwareZone found an ASUS ProArt P16 faster than an M5 Max MacBook Pro at prompt processing but about half the speed at token generation, and roughly equal on a 50-page document overall.
• Images and video: diffusion models are much more compute-bound than chatbots, so FP4 and tensor-core throughput matter more.
• Many users at once: batching several requests reuses each weight read for many tokens, which shifts the bottleneck towards compute. See is the GR1X powerful enough for multiple users.

Does the GR1X's 140W cooling change the numbers?

For token generation, probably not much. Every RTX Spark machine shares the same memory bandwidth, so a desktop running at 140W shouldn't generate tokens noticeably faster than a laptop at 45–80W. Where higher sustained power may help is the compute-heavy work above: prefill, image generation, rendering and gaming. That's our reading of the hardware, not a measured result; no GR1X benchmarks had been published at the time of writing. We look at this in does better cooling mean faster AI.

How should you judge AI performance before buying?

• Ask for tokens per second on a named model, quantisation and context length, not TOPS or petaflops.
• Check memory bandwidth and how much memory the GPU can actually use (about 64GB by default on 128GB RTX Spark laptops, adjustable to about 96GB).
• Prefer mixture-of-experts models such as gpt-oss-120b on bandwidth-limited machines; they read far less per token than dense models of the same size.
• Look at prefill speed if your work involves long documents.

Our guide to NVIDIA's 120B and 200B model claims and the 128GB memory trap go further.

Buying in Singapore

The ASUS ProArt GR1X is listed at SourceIT as price on request. A SourceIT estimate based on comparable RTX Spark devices (not an official ASUS price) is about S$6,200–7,200 for 64GB/1TB and S$9,000–10,800 for 128GB/2TB, including GST. To compare the same chip in a laptop now, the ASUS ProArt P16 128GB/1TB is S$10,773.20, and the NVIDIA DGX Spark 4TB 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 GR1X's 1 petaflop the same as the DGX Spark's? Both are headline FP4 figures of 1 petaflop. NVIDIA footnotes the DGX Spark figure as theoretical with sparsity; its RTX Spark page doesn't carry that footnote.

How many tokens per second will the GR1X generate? No GR1X benchmark had been published at the time of writing. A same-chip laptop ran gpt-oss-120b at about 63 tokens per second. Dense 70B models at 4-bit are limited to about 7 tokens per second by memory bandwidth.

Why is a Mac Studio faster at generating text without a petaflop rating? Because generation speed depends mainly on memory bandwidth. The M5 Ultra's 1.2 TB/s is about four times the N1X's roughly 300 GB/s.

Does a higher petaflop rating ever help? Yes, for compute-heavy work: reading long prompts, generating images and video, and serving several users at once.

What should I compare instead of petaflops? Memory bandwidth, usable GPU memory, and published tokens per second for a named model, quantisation and context length.