SemiAnalysis: NVIDIA and Inferact's first vLLM submission on Rubin NVL72 is state-of-the-art

SemiAnalysis says NVIDIA and Inferact have submitted an initial vLLM result on Rubin NVL72, less than three weeks after the MiniMax M3.1 Flash launch. SemiAnalysis calls the submission state-of-the-art and says it delivers 6.5x better performance per total cost of ownership than GB300 NVL72.

According to SemiAnalysis, the submission uses the new locality-aware MoE feature in CUDA 13.4, which it says reduces latency and raises realized HBM bandwidth by about 1.2x. It describes the work as done in partnership with the vLLM project.

The figures are SemiAnalysis's own claims.