Model Acceleration
Same model. 1.54× faster.
Model acceleration is the discipline of making a model faster without changing what it is: attention, caching, decoding, and quantisation work around the weights. The case study is MiniMax H3 — a 33B video model — accelerated on a single DGX Spark by following NVIDIA's Sol-Engine approach and porting it in-house, packaged as three ComfyUI node packages with an executable recipe and the full raw benchmark ladder.
The headline
A 5-second clip — 864×480, 124 frames, video with stereo audio — rendered stock and then with the acceleration stack, same model, same box:
| End-to-end clip time | 312.7 s → 202.5 s1.54× faster, measured |
|---|---|
| Per-step time | 14.40 s → 8.39 s1.72× per diffusion step |
The stack
Keel H3 Turbo is the model-agnostic acceleration engine; MiniMax H3 is the current supported model and the case study below. Explore the source on GitHub.
| Sol-Attn sparse attention | 1.32×Sol-Engine-style sparse attention, ported in-house — plus an INT8 QK/TMA variant |
|---|---|
| FirstBlockCache | 1.48×First-block features cached across steps instead of recomputed |
| Batched VAE tile decode | 1.66×Tiled VAE decoding, batched |
| AdaLN precompute | —Conditioning computed once, not per step |
| Pruned FP8 checkpoints | 35.7 GB → 21.0 GBBF16 checkpoint set pruned to an FP8 working set |
| NVFP4/AWQ text encoder | —The Qwen3-VL-32B encoder quantised for the Spark |
| GB10 spin-wait patch | —Platform fix for the GB10's busy-wait behaviour |
Production proof
The stack is not a benchmark toy — it is the engine under a short film now in production. The film's first keeper shot (1024×576, 192 frames, 20 steps) rendered in 710.15 seconds on this stack, with a signed receipt for the take. Takes are generated against locked stills, logged, and marked failed when they miss the bar — the same measured discipline as everything else on this site.