vLLM reports DeepSeek-V4.1-Flash throughput on Vera Rubin NVL72Machine translation
vLLM 报告 Vera Rubin NVL72 上的 DeepSeek-V4.1-Flash 吞吐表现
On a single NVIDIA Vera Rubin NVL72 rack, vLLM serves DeepSeek-V4.1-Flash in SemiAnalysis AgentX at 1.7 million tokens per second per GPU, with a user-side P90 of 387 tokens per second. The reported throughput is 58 times that of GB300 NVL72 at the same interactivity.
This source permits summary display only.
Read at the original source