跳到正文

NVIDIA Vera Rubin 首次亮相并在 MLPerf 训练与智能体推理基准测试中取得领先成绩

NVIDIA Blog当事人· Shruti Koparkar·· 8 小时前精选NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance
打开原文
原文
图片来自 NVIDIA Blog
谁
NVIDIA
时间
引文
In its MLPerf benchmarks debut, with results released today, the NVIDIA Vera Rubin platform delivers leading performance across training, post-training and agentic inference.

发生了什么

NVIDIA 表示,其 Vera Rubin 平台在首次亮相中横扫了每一项 MLPerf Training v6.1 基准测试,并在 MLPerf Endpoints 智能体基准测试中全面领先。在 DeepSeek-V3 训练中,该平台从 256 颗 GPU 扩展至 1,024 颗 GPU 时仍保持了约 97% 的效率;一项暂定 Kimi K3 结果显示,在每位用户每秒 50 个 token 的相同最低交互速度下,其每瓦吞吐量达到 GB300 NVL72 的 10 倍以上。Vera Rubin 的 Endpoints 测试结果目前正等待 MLCommons 验证。

推荐理由

Vera Rubin 在 DeepSeek-V3 训练中从 256 颗 GPU 扩展至 1,024 颗 GPU 时保持了约 97% 的效率,同时其暂定 Kimi K3 结果显示,在相同响应速度下,其每瓦吞吐量达到 GB300 的 10 倍以上。

接下来读什么

查看这件事的完整进展NVIDIA Reports Vera Rubin Debut Results in MLPerf Training and Endpoints

这件事到目前为止

NVIDIA said its Vera Rubin platform swept every MLPerf Training v6.1 benchmark and led across MLPerf Endpoints agentic benchmarks in its debut, though its Endpoints results remain pending MLCommons Association verification. On DeepSeek-V3 training, NVIDIA reported that Vera Rubin scaled from 256 to 1,024 GPUs with about 97% efficiency. For MLPerf Endpoints v1.0, a provisional Kimi K3 result showed more than 10x the throughput per watt of GB300 NVL72 at a shared minimum interactivity of 50 tokens per second per user. NVIDIA noted that the provisional Vera Rubin results are subject to change pending peer review.

这条收录于 11月17日 AI 日报看这一期日报的全部内容