NVIDIA Sol Engine accelerates MiniMax H3 to 14.93s on GB200

NVIDIA's Sol Engine splits generation into a 4-step low-res H3 draft and a 3-step LTX refinement pass with Sol-Attn, cutting 768p latency on a single GB200 from 414s to 14.93s. A separate Sol-Attn SM89 CuTe DSL kernel brings 4.44× end-to-end speedup on RTX 4090.
How this story unfolded
8 days · 0 reports · 3 community posts · from Aug 17
MiniMax by email
Get an email when MiniMax has news
No news that day, no email.
More stories today
- Abliterlitics compares 12 abliterated Gemma 4 12B variants
- ElevenLabs launches Composer section-by-section song editor in ElevenMusic
- Perplexity CEO thanks Nvidia for DGX Spark research support
- LangChain cookbook builds due-diligence agent with Deep Agents and Parallel
- LangSmith and LangChain OSS help meet EU AI Act requirements