ProductsNov 16, 2026, 11:00 AM EST

NVIDIA Vera Rubin NVL72 Tops MLPerf Endpoints Agentic Inference Results

NVIDIA published provisional MLPerf Endpoints v1.0 results for its Vera Rubin NVL72 rack-scale system, setting it against verified GB300 NVL72 figures across the Kimi K3, DeepSeek-V4.1-Flash and Qwen3.6-35B-A3B models. At a shared minimum of 50 tokens per second per user of end-to-end interactivity, the company reports more than 10x the throughput per provisioned kilowatt, with 5x to 10x gains for individual models and 3x to 5x higher interactivity at matched concurrency. The Rubin results are provisional and still pending MLCommons verification. Rubin GPUs pair HBM4 memory, Tensor Cores and sixth-generation NVLink; NVIDIA also credits prefix caching, state reuse and Dynamo routing for the serving gains.

Why it matters

Agentic sessions pile up long conversation histories, so serving capacity at a fixed level of responsiveness determines how many users a deployment can support. If the provisional figures hold up in verification, the same power budget could serve substantially more agent traffic — but NVIDIA's numbers have not yet been reviewed by MLCommons.

Sources

Written by AIToolMaps from the source above and checked against it sentence by sentence.

Comments

No comments yet