NVIDIA BioNeMo recipe boosts MoE biological model training up to 2.21x on eight B200s
AllThingsApx · x · 2026-10-01
NVIDIA's technical blog details an efficient MoE training recipe for biological foundation models using BioNeMo and Transformer Engine:
- GroupedLinear: submits multiple expert GEMMs as one grouped operation, removing kernel-launch overhead from per-expert Python loops.
- MXFP8: block-scaled 8-bit precision cuts memory vs BF16 with Blackwell hardware acceleration.
- Kernel fusion: the Sequential API fuses GroupedLinear, ScaledSwiGLU, and routing-weight scaling into a single kernel, avoiding intermediate materialization.
- Benchmark: up to 2.21x throughput over the Hugging Face baseline on eight B200 GPUs. A Mixtral Native Transformer Engine recipe is available to try.
Related event: NVIDIA BioNeMo Speeds Up MoE Training Up to 2.21x Over HF Baseline(3 posts)→
More from Infra
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Stratechery's Fictional Obituary: Imagining Jensen Huang's 2026 Exit From Nvidia — Gregory_C_Allen · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16
- Colossus 2's 1.2M B200s could generate ~$400B/year in tokens at DeepSeek Flash rates — bookwormengr · 2026-11-16
- BIS issues rule blocking remote access to US AI chips via third-country data centers — teortaxesTex · 2026-11-16
- Qdrant 1.20.1: 50% fewer cold-storage reads, ~2x faster quantization, agent-ready setup — qdrant_engine · 2026-11-16