AI News

Trending story

NVIDIA Announces Preliminary Results for Vera Rubin NVL72 on MLPerf Endpoints v1.0

1 reports1 story updatesUpdated Nov 17 00:00

Understand this story

AI overview

On November 17, 2026, NVIDIA announced provisional results for Vera Rubin NVL72 on MLPerf Endpoints v1.0. An article published by NVIDIA Developer · Generative AI states that, under the condition of end-to-end interactivity of no less than 50 tokens/s per user, Vera Rubin NVL72's throughput per kilowatt exceeds that of GB300 NVL72 by ten times. The article further explains that on the three models Kimi K3, DeepSeek-V4.1-Flash, and Qwen3.6-35B-A3B, Vera Rubin NVL72's throughput per kilowatt is 5 to 10 times that of GB300 NVL72, and at the same concurrency, end-to-end interactivity is 3 to 5 times higher. The article also introduces optimizations such as prefix caching, state reuse, NVIDIA Dynamo routing, disaggregated serving, as well as the DeepSeek MegaMoE kernel and Qwen multi-token prediction. There are currently no subsequent development facts.

AI-generated from public reports

Latest development · Report

NVIDIA Vera Rubin NVL72 on MLPerf Endpoints…

Developments

1 confirmed developments

  1. NVIDIA announces preliminary results for Vera Rubin NVL72 on MLPerf Endpoints v1.0

    Source:NVIDIA Developer · Generative AI

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 17

  1. NVIDIA Developer · Generative AIFirst-partySelected

    NVIDIA Vera Rubin NVL72 on MLPerf Endpoints…

    NVIDIA announced the provisional MLPerf Endpoints v1.0 results: under the condition of per-user end-to-end interactivity of no less than 50 tokens/s, Vera Rubin NVL72 achieves more than ten times the per-kilowatt throughput of GB300 NVL72. On the three models Kimi K3, DeepSeek-V4.1-Flash, and Qwen3.6-35B-A3B, Vera Rubin NVL72's per-kilowatt throughput is 5 to 10 times that of GB300 NVL72, and at the same concurrency its end-to-end interactivity is 3 to 5 times higher. The article also introduces optimizations such as prefix caching, state reuse, NVIDIA Dynamo routing, disaggregated serving, as well as the DeepSeek MegaMoE kernel and Qwen multi-token prediction.

Related stories

  1. Related storiesNVIDIA Vera Rubin NVL72 Makes Its Debut in MLPerf Training v6.1
  2. Related storiesGoogle launches Gemini 4 Hydrogen on Antigravity
  3. Related storiesMeta releases open-weight model Muse Spark Open
  4. Related storiesNvidia releases Nemotron 3.5 Super VL and Bolt models