MiniMax H3 Benchmark: Native implementation 30%+ faster

Mattnix · reddit · 2026-08-15

A benchmark of MiniMax H3 INT8 on the AMD Radeon RX 9070 XT (16GB VRAM) shows that ComfyUI's native INT8 implementation is approximately 31% faster at 1MP resolution than PatientX's INT8-Fast-ROCM implementation (native took 32 mins). Using identical models, prompts, and steps, the results contradict documentation suggestions, implying existing optimization advice may be outdated or RDNA4 behaves differently.

Related event: MiniMax H3 INT8 Benchmark: ComfyUI Native Implementation Beats Fast-ROCM by 45% on RX 9070 XT(2 posts)→

Original post →

More from Infra

Infra channel →