On 10 July AMD published benchmarks on its blog for generative visual models on the accelerator MI350X. The numbers are impressive, but the baseline is an unoptimized reference implementation - not a previous-generation chip, not NVIDIA. The tests are AMD's own, with no independent verification.
- By AMD's numbers: FLUX.1-dev generates an image 2.25x faster on MI350X through the software SGLang Diffusion.
- By AMD's numbers: Z-Image-Turbo generates 6.29x faster, and editing with Qwen-Image-Edit is 5.78x faster.
- The baseline is the standard unoptimized HuggingFace Diffusers implementation, not a previous AMD chip and not NVIDIA.
Every vendor benchmark with an 'X times faster' number hides one question you need to ask before you believe it: faster than what. On 10 July AMD published results on its blog for the MI350X accelerator in image generation. At first glance it sounds strong. Up close, the number answers a completely different question.
I'll believe it when I see how the MI350X holds up against an NVIDIA chip, on the same software. That's exactly the comparison that's missing. AMD is measuring its own optimized software against a baseline implementation nobody runs in production in that shape. The 6x answers how much AMD gains from its own optimization - not whether it's faster than the competition.
Don't read this as proof the MI350X beats NVIDIA - the post doesn't claim that, and we shouldn't put words in its mouth. The more important signal is that AMD is putting serious work into the software layer, and that's exactly where it's been weak for years. Its hardware hasn't been the problem for a while now. Whether this turns into an advantage will show up in independent tests down the line.