AMD: Instinct MI355X outperforms NVIDIA B200 on ComfyUI workflows with PyTorch optimizations in ROCm 7.2.0

AMD has released benchmarks showing that the Instinct MI355X outperforms NVIDIA’s B200 across three ComfyUI generative workflows through PyTorch attention optimizations for the CDNA4 architecture in ROCm 7.2.0. Results are distributed via a Docker image that any user can run.

Results by workflow

The numbers are clear:

• Text-to-Video (Wan2.2): MI355X achieves a 1.439× speedup with a runtime of 116.91 seconds versus 168.28 s on B200.
• Text-to-Image (FLUX.1-dev): 1.416× faster, 24.77 s vs. 35.09 s.
• 3D Generation (Hunyuan3D v2.1): 1.201× speedup, 21.51 s vs. 25.84 s.

CDNA4 hardware

MI355X brings 256 compute units, 288 GB of HBM3e memory, and memory bandwidth of 8 TB/s. These specifications favor the attention-heavy operations typical of diffusion models — large memory holds intermediate representations of high-resolution images and video frames without tiling, and bandwidth reduces memory wait time.

Optimizations in ROCm 7.2.0

Performance does not come from hardware alone. The key software additions are:

• AOTriton gfx950 kernel support — native attention acceleration via the ahead-of-time Triton compiler for CDNA4.
• Occupancy tuning — reduced warp count for better efficiency.
• hipBLASLt GEMM optimization — tuned kernels for FP8, BF16, and FP16, the dominant types in diffusion and transformer workloads.
• Pipelining and ThinLTO compiler optimizations.

What this means for the AI hardware market

Three production scenarios — video, image, 3D — cover the majority of generative ComfyUI use cases. That AMD shows consistent numbers across all three signals that the CUDA/ROCm gap, long a structural advantage for NVIDIA, is narrowing depending on the software maturity of individual frameworks. For ComfyUI users, AMD is now a legitimate option — at least at the benchmark level.