硬體
ASUS Mini Desktop Cuts Image Generation Time by Nearly One-Fifth, with More Limited Gains for Text
The ProArt GR1X offers up to 128GB of unified memory, and the review lists a starting price of $4,499. For local AI users weighing the investment, model capacity, benchmark scores, and actual wait times need to be considered separately.

Users preparing to buy a dedicated computer for running local AI have another desktop option with up to 128GB of memory: the ASUS ProArt GR1X. After comparing it with a laptop on the same platform, Windows Central measured text generation that was about 4% to 8% faster for some models, while SDXL image generation took nearly 18% less time. The review’s “up to 21.9% improvement” refers to a score increase and cannot be applied directly to every generation workload. Original review
Generation speed and composite scores are different measures
Windows Central’s Zac Bowden and Cale Hunt used the Procyon benchmark to compare the GR1X with the Surface Laptop Ultra. Both use the RTX Spark N1X, and the primary comparison is between 128GB configurations. The review lists maximum TDPs of 140W for the desktop and 80W for the laptop. TDP is a thermal design power metric; it should not be treated as measured power consumption at the wall. Test configurations
Phi-3.5-mini’s output increased from 72.21 to 77.95 tokens per second, which works out to about 7.9% faster; Mistral 7B increased from 53.07 to 55.4, or about 4.4% faster. A token is a unit the model uses to process text, and it does not correspond directly to a Chinese character count. Both of these sustained output gains are smaller than the composite score increase cited in the review. Results table
UL’s official documentation explains that the Procyon text score combines time to first token with subsequent output speed. For long-form generation, sustained output rate matters; for an interactive assistant, first-response latency matters too. A single score cannot fully capture both experiences. UL scoring explanation
About four and a half seconds less waiting per image
The difference is more direct for Stable Diffusion XL: each image took 20.6 seconds on the 128GB desktop, compared with 25.05 seconds on the laptop—about 4.45 seconds less, or 17.8% shorter. The review’s reported maximum 21.9% score improvement corresponds to the slightly faster 64GB desktop result. Image generation test
An earlier laptop review identified this SDXL test as FP16, or half-precision floating-point computation. The desktop article does not fully specify the precision, inference engine, or generation settings, so these results represent that publication’s test procedure. They do not establish that a custom ComfyUI workflow would see the same gains. Both reviews are from the same publication, so they are not two independent validations. Laptop test baseline
More memory does not necessarily mean faster text generation
ASUS has announced 64GB and 128GB configurations. The announcement page is dated November 19, while the announcement text gives November 18. Unified memory means the CPU and GPU share a memory pool. ASUS also claims support for 120B models—those with 120 billion parameters—but that capacity claim is not evidence of large-model speed. ASUS availability announcement
Windows Central lists a starting price of $4,499, while ASUS notes that pricing and availability vary by region. Based on the results above, people who regularly generate images in batches are more likely to save meaningful time. Those primarily running small text models should not decide to buy based only on a benchmark score increase of around one-fifth. Readers who need large models still lack throughput data for those models from this review. Pricing information, regional availability details
Sources
- ASUS Announces Availability of ProArt GR1X
- ASUS ProArt GR1X review: The first RTX Spark mini PCs have landed, but more performance is offset by a significantly higher price — I tested to see if it's worth it
- Absolutely everything I learned while testing RTX Spark for a week: It's a beast, but not at everything
- How is the Procyon AI Text Generation Benchmark score calculated?