NVIDIA showed how to speed up MoE training for biological foundation models by pairing kernel-level tricks with Blackwell-only low-precision math.
NVIDIA published an efficiency recipe for training Mixture-of-Experts (MoE) foundation models in the life sciences, built on its BioNeMo framework. MoE cuts compute versus dense transformers, but fragmented per-expert computation lowers GPU utilization, and routing strains communication and memory. The recipe batches expert computations with a "GroupedLinear" operation, uses 8-bit "MXFP8" precision, and fuses the matmuls into one kernel, hardware-accelerated specifically on Blackwell-generation GPUs.
On eight B200 GPUs, the recipe reached 2.21x the throughput of a Hugging Face baseline (9,050 tokens per second per GPU versus 4,096). It suggests efficiency techniques refined on general-purpose LLM training carry over to a domain, biological foundation models, whose data is far less uniform. The figures come from one specific setup, eight Blackwell-generation GPUs, so the same multiple isn't guaranteed at smaller scale or with a different model.
Whether this recipe extends beyond BioNeMo to other life-science foundation models, or to MoE models in unrelated domains, is the open question. So is whether Blackwell-only formats like MXFP8 become a pattern other hardware makers can reproduce, or stay a one-vendor advantage.