One attention node makes local MiniMax-H3 video generation 2.5x faster on AMD Radeon 8060S iGPU
ShamanFlamingoFR · reddit · 2026-09-21
Three weeks after running MiniMax-H3 video generation locally on a Ryzen AI Max+ 395 (Radeon 8060S iGPU, 128GB unified memory), the author reports per-evaluation time dropping from 45.7s to 18.3s and end-to-end cold renders from 521s to 206s — same machine, same graph, same model. The lever: ComfyUI's ModelAttentionBackend("comfy kitchen attention") node finally has a working kernel.
Key details
- comfy-kitchen 0.2.34 (pinned by ComfyUI 0.36.0) adds solattn/solattnchunked with gfx1150–1153 support — an officially supported path, not luck. Check the startup log for solattn, or the node silently does nothing.
- No quality trade: 31.6 dB PSNR, 0.995 audio correlation. A 3-block 15s clip render went from 31m40s to 13m11s (2.4x).
- The pytorch attention option still crashes on fp8 trunks via HIP (SDPA path not ready on gfx1151).
- The Windows DynamicVRAM crash is fixed with comfy-aimdo 0.5.3; torch fragmentation collapsed from 21.1GB to 2.0GB.
Revised conclusion: a 2.5x speedup from just swapping attention at unchanged clocks means the earlier "bandwidth-bound" verdict was wrong — the bottleneck was split attention's intermediate VRAM round-trips, which the fused INT8 kernel eliminates.
More from Fun
- ICLR 2027 and beyond: is it the author or the agent who's doing the work? — MuCai7 · 2026-09-21
- World's first human vs. humanoid robot fight held, a 6ft robot steps into the ring — Admirable-Cell-2658 · 2026-09-21
- "Personal Superintelligence makes zero sense": dev rants AGI/ASI diluted into buzzwords — menhguin · 2026-09-21
- Maze test: a pure random generator beats AI model Jev, which loops 2,295 steps in a corner — karminski3 · 2026-09-21
- The 'romance novelist Caroline Ellison' meme, explained — verdakorz · 2026-09-21
- s-risk redefined: the odds your late-night SeaTac flight lands at terminal S — EigenGender · 2026-09-21