Accelerating ComfyUI Workflows on AMD Instinct™ MI355X GPUs with ROCm

2026-05-11 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI Hardware Acceleration · Depth: Advanced, long

Summary

AMD's recent blog post, dated May 11, 2026, details how ComfyUI generative AI workflows are accelerated on AMD Instinct MI355X GPUs using ROCm 7.2.0. The article benchmarks three production-relevant ComfyUI workflows—Wan2.2 (Text-to-Video, 1280x1280), FLUX.1-dev (Text-to-Image, 2560x2560), and Hunyuan3D v2.1 (Image-to-3D, 4096)—against an NVIDIA B200 GPU. The MI355X, built on the CDNA4 (gfx950) architecture, achieved up to 1.439x faster end-to-end execution. This performance advantage is attributed to PyTorch Attention optimizations for gfx950, including AOTriton kernel support, occupancy tuning, pipelining, hipBLASLt GEMM tuning, and ThinLTO compiler optimizations, all integrated into ROCm 7.2.0 and available via a pre-built Docker image.

Key takeaway

For AI Engineers evaluating hardware for generative AI workloads, the AMD Instinct MI355X with ROCm 7.2.0 offers a compelling performance advantage, demonstrating up to 1.439x faster execution in ComfyUI workflows compared to the NVIDIA B200. You should consider leveraging the pre-built Docker images to quickly deploy and benefit from the optimized PyTorch Attention and GEMM kernels, potentially reducing inference times for compute-intensive diffusion models.

Key insights

AMD Instinct MI355X GPUs significantly accelerate ComfyUI generative AI workflows over NVIDIA B200.

Principles

• Hardware-software co-optimization drives performance.
• Attention mechanisms are critical for diffusion model speed.

Method

Benchmarking involved three attention-heavy ComfyUI workflows on AMD Instinct MI355X (ROCm 7.2.0) and NVIDIA B200 (CUDA 12.x), measuring end-to-end execution time using default parameters.

In practice

• Use pre-built Docker images for optimized ComfyUI setup.
• Explore xDiT for multi-GPU ComfyUI inference scaling.

Topics

• AMD Instinct MI355X
• ComfyUI Workflows
• ROCm 7.2.0
• PyTorch Attention Optimizations
• Generative AI Benchmarking

Code references

• Comfy-Org/ComfyUI
• Wan-Video/Wan2.2
• black-forest-labs/flux
• Tencent-Hunyuan/Hunyuan3D-2.1
• xdit-project/xDiT

Best for: AI Hardware Engineer, Machine Learning Engineer, AI Engineer

Related on AIssential

• AMD Instinct™ GPUs MLPerf Inference v6.0 Submission — AMD's MI355X GPU and ROCm stack deliver competitive MLPerf Inference v6.0 results, especially for LLMs and multi-node setups.
• Taking on CUDA With ROCm: ‘One Step After Another’ — AMD's ROCm software stack is rapidly maturing, leveraging open-source tools and community engagement to challenge Nvidia's CUDA dominance.
• Open-Source AI Tool Upgrades Speed Up LLM and Diffusion Models on NVIDIA RTX PCs — NVIDIA's CES 2026 announcements significantly boost AI PC developer capabilities through framework optimizations and new model releases.
• Efficient and Portable 3D Explorable World Generation on AMD GPUs — Optimizing 3D world generation on AMD GPUs requires kernel portability and pipeline efficiency.
• Dell/AMD partnership: Three insights you may have missed from theCUBE’s coverage of Dell Technologies World — Dell and AMD's partnership enables hybrid AI architectures, leveraging PCIe for air-cooled on-prem inferencing and driving data center consolidation.
• ROCm with PyTorch and PyTorch Lightning seems to still suck for research [D] — ROCm remains challenging for ML research due to instability and NaN issues with less common PyTorch codebases.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.