ai& cuts MiniMax H3 video generation from 229s to 19s across NVIDIA, AMD, Tenstorrent
DavidBennett__ · x · 2026-10-24
Japanese heterogeneous-inference provider ai& published a deep technical blog on optimizing the H3 video model across three architectures: NVIDIA B300, AMD MI355X, and Tenstorrent Blackhole.
- Caching: per-refinement-step time cut from 2.53s to 0.05s
- Step distillation: 15-second video generation dropped from 229s to 19s
- Baselines: 8×B300 36 PFLOPS FP8 / 2.1TB HBM3E; 8×MI355X 40.27 PFLOPS / 2.3TB; Blackhole Galaxy 23 PFLOPS single, 92 PFLOPS 4-node, 8-10kW each
The post walks through compute/memory/bandwidth tradeoffs per hardware and the quality-vs-cost costs of each speedup — a solid reference for diffusion video model inference engineering.
More from Infra
- AI credit risk hits new highs as Oracle 5Y CDS spikes to ~300 — AccBalanced · 2026-10-24
- Nebius is hoarding GPUs to sell at scarcity prices — funded by a $3.5B 364-day bridge loan — tengyanAI · 2026-10-24
- Spark-X2.6-4B GGUF quantized model trends on Hugging Face under Apache 2.0 — XHToken · 2026-10-24
- High Refractive Index Glass Hits 2.1 With 97% Transmittance, Raising Eyebrows for AR Optics — jwt0625 · 2026-10-24
- EPFL demos ~100GHz-spacing TFLT WDM modulator chip; expert doubts CPO fit — jwt0625 · 2026-10-24
- Reddit deep-dive: next-gen AI via memristors, spiking nets, and 1% sparse routing — Robert__Sinclair · 2026-10-24