MoE · Mixture of Experts
A sparse model that activates only a subset of expert sub-networks per token, cutting compute per token at scale.
Current numbers
~1/4 the number (~75% fewer, ~4×)Vera Rubin GPUs to train an MoE model vs Blackwell (training metric)
~10xtokens/watt advantage of Blackwell-class over Hopper on MoE inference — the power-limited lever
~7x / ~10xDynamo + wide-EP MoE throughput on GB200 NVL72 vs B200 (Dynamo 1.0 GA at GTC 2026); NIXL+GPUDirect Storage prefill speedup for long context