Skip to content

Performance

Jump to bottom
AlexHe edited this page Sep 3, 2026 · 10 revisions

Performance

Headline numbers for main. Three HIP ISAs share one tree; build with an explicit HIP_ARCH (Getting started). Wall time moves with page-cache state; treat E2E as a band. Denoise GPU is the stable GPU figure.

In-tree copy: docs/PERFORMANCE.md.

Current (main, 2026-09-03)

Preset gfx1151 gfx90a gfx942
fox-s2 E2E ~85–90 s (I/O) ~10.8 s ~16 s
fox-fast E2E ~2 min (I/O) ~18 s ~12 s
15 s cinematic E2E 40 min 46 s 12 min 11 s ~2.5 min

gfx1151 v0.9.0 fox-s2 md5 1731f95c4aa582597cf83d57f46b8f9e. On current main default 512 px VAE tiles, fox-s2 is 34507f072c5cabbde6592b3f70b8fa35.

Halo 15 s quality path: denoise 36 min 38 s, VAE 174 s (2×1 @ 480 px), DiT peak 27.9 GiB. Opt-in --token-reduction 15 s is still 28.2 min (measured 2026-09-02 vs then-45.0 min). Details: Long video.

# fox-s2 — HIP A/B smoke (gfx1151 md5 gate)
./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. fox-s2 / fox-fast are complete ~0.9 s MP4s, not truncated previews. Showcase fox uses --layers 50 --reuse 1 and is not the fox-fast scoreboard.

Long T2VA: Long video. gfx90a DiT is BF16 by default (H3_INT8_MLP=1 restores INT8).

Optional --token-reduction (off by default; same CLI as h3-spark.c) pairs middle-block video tokens. Do not replace the tagged quality-path row (45.0 min / 12 min 33 s) with these numbers.

quality path --token-reduction
gfx1151 15 s E2E 45.0 min 28.2 min (−37%); perf-runs/TOKEN_REDUCTION.md
gfx90a 15 s E2E 12 min 33 s 8 min 21 s (−34%); perf-mi210/TOKEN_REDUCTION.md

gfx1151 history (v0.9.0)

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
v0.10.0 same on gfx1151 ~18 s on gfx90a Dual ISA; MI210 MFMA flash

h3-hip.c

Project page · Repository · v0.11.0

Clone this wiki locally