vLLM reports 10.2x MiniMax-M3.1-Flash throughput gain on NVIDIA Vera Rubin

1. 10:40 PM
   SemiAnalysis@SemiAnalysis_X

   NVIDIA and Inferact submit initial vLLM Rubin NVL72 benchmark result

   AINVIDIA and Inferact have submitted an initial vLLM benchmark on Rubin NVL72 that the post calls state-of-the-art, with 6.5x better performance per TCO than GB300 NVL72. The submission uses a new locality-aware MoE feature in CUDA 13.4, which the post says reduces latency and increases realized HBM bandwidth by about 1.2x.

   
2. 6:25 PM
   vLLM@vllm_projectOfficial

   vLLM reports further Rubin performance gains on MiniMax M3.1 model

   AIvLLM says its engine optimizations and Rubin-optimized kernels, including MiniMax Sparse Attention and decode-specialized kernels, have pushed performance further since its M3.1 blog post. The gains build on earlier Rubin work covered in the M3.1 and Vera Rubin preview posts. vLLM says it is "just getting started" on further optimizations.

3. 6:25 PM
   vLLM@vllm_projectOfficial

   vLLM reports 10.2x token throughput for MiniMax-M3.1-Flash on NVIDIA Vera Rubin

   AIvLLM says MiniMax-M3.1-Flash runs 10.2x faster per chip on NVIDIA Vera Rubin NVL72 than on GB300 NVL72 in SemiAnalysis AgentX, at 546 P90 tokens per second per user. The gain is 6.5x at 376 P90 tokens per second per user, and up to 1,205 P90 tokens per second per user at low concurrency.

   

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.