Bypassing PyTorch with HIP kernels makes MiniMax H3 up to 3x faster on AMD R9700
Apprehensive_Sky892 · reddit · 2026-11-19
A developer released a standalone app (v1.2) running MiniMax H3 on AMD AI Pro R9700 up to 3x faster than the same config on ComfyUI, by bypassing PyTorch/Triton/ComfyUI entirely and running directly on AMD's HIP C++ runtime within ROCm.
Benchmarks (5s clip, Turbo4):
- 864x480, 1 image: 67.4s → 30.4s
- 864x480, 3 images: 91.2s → 33.4s
- 864x480, 9 images: 167.5s → 44.3s
- 1344x768, 9 images: 278.7s → 114.1s
Gains grow with more reference images because the stock engine spends over half its time encoding them. Limits: 9.4s of 768p video with 9 images, no audio-only references, 50-step Base preset still on v1.1.0. Shows AMD consumer cards still have big headroom for video model inference.
More from Infra
- KAIST Proposes PROLA: Low-Rank Adaptation for Spatiotemporal Weather Downscaling — MLVLab · 2026-11-19
- Ex-DeepSeek core member open-sources Linger, a local AI agent claiming 5x faster inference — Aiden_Tech_Ai · 2026-11-19
- China opens first overseas-targeted AI cluster on Huawei Ascend chips — pstAsiatech · 2026-11-19
- Tencent says US chip curbs 'unfortunate', may slow AI training short-term — pstAsiatech · 2026-11-19
- Samsung dismisses AI bubble fears: neoclouds are just 8% of the market, chip shortage to persist — BenBajarin · 2026-11-19
- Samsung halts China AI chip foundry work under new BIS rules, hitting Cambricon's advanced-node pipeline — teortaxesTex · 2026-11-19