Performance

Performance (gfx1151)

Headline numbers for tagged releases, reproduced on an AMD Ryzen AI MAX+ 395 / Radeon 8060S with the official MiniMax-H3 checkpoint.

Wall time moves with page-cache state; treat E2E as a band. Denoise GPU (HIP events) is the stable GPU figure.

In-tree copy: docs/PERFORMANCE.md. Session logs under docs/perf/ are not user-facing.

v0.9.0 (2026-08-26)

Preset	Command knobs	E2E	Denoise GPU
fox-s2	512² · 22 frames · --steps 2 --layers 35 --reuse 1	83–87 s	6.3–6.5 s
fox-fast	512² · 22 frames · --steps 20 --layers 45 --reuse 2	95 s	28 s (11 evals)

fox-s2 output md5 1731f95c4aa582597cf83d57f46b8f9e on that tree.

# fox-s2 — HIP A/B smoke (bit-identical check)
./h3 --profile -d /path/to/MiniMax-H3 \
  -p "A red fox walks through fresh snow." \
  --width 512 --height 512 --frames 22 --steps 2 --layers 35 --reuse 1 \
  -o outputs/fox-s2.mp4

# fox-fast — upstream tutorial “first fast video”, without --show
./h3 --profile -d /path/to/MiniMax-H3 \
  -p "A red fox walks through fresh snow in a pine forest. Medium tracking shot, natural winter light, realistic fur, soft footsteps and wind." \
  --width 512 --height 512 --frames 22 --steps 20 --layers 45 --reuse 2 \
  -o outputs/fox-fast.mp4

--reuse 2 on 20 steps yields 11 DiT evaluations, not 20. Showcase fox uses --layers 50 --reuse 1 (20 evals) and is not the fox-fast scoreboard.

First invocation pays ~107 GiB of T2VA weight reads. Remaining E2E is mostly NVMe I/O; denoise is a small slice of fox-s2 and about a third of fox-fast.

High-level history

When	fox-s2 E2E	fox-fast E2E	What landed
2026-08-18	~257 s	—	First HIP profile
2026-08-19	~117 s	—	INT8 DiT, faster load I/O
2026-08-22	—	~213 s	First fox-fast measure
v0.9.0	83–87 s	95 s	WMMA kernels; weights in VRAM carveout; AdaLN / VAE load overlap

Upstream antirez/h3.c publishes denoise wall on an M5 Max, not T2VA end-to-end. Same fox-fast knobs: M5 16.7 s denoise wall vs HIP ~35 s. That ratio mixes two GPUs and two memory systems; it is not a port-quality score.