MLPerf Training v6.1 发布:GB300 NVL72 上 GPT-OSS-20B 训练仅 19.697 分钟
想搞清楚 H100、Blackwell 和 Blackwell Ultra 差多少的可以看这份榜单:同一套 64 卡 GB300 系统把 GPT-OSS-20B 训练压到不到 20 分钟,比 H100 快近 3 倍。计划采购 GPU 或做大模型训练的团队值得拿这些数字做预算参考。
MLPerf Training v6.1 榜单出炉。Lambda 在单套 NVIDIA GB300 NVL72 系统(64 张 Blackwell Ultra GPU)上,将 GPT-OSS-20B 训练用时压到 19.697 分钟,比上一轮 64 张 H100 的成绩快 2.86 倍,比 64 张 Blackwell GPU(CoreWeave)快 1.37 倍。
Lambda 还首次跑了一个多模态基准:在同一套系统上非官方复现 Flux.1 训练,用时 26.36 分钟,比上一轮同硬件最快成绩快 1.50 倍。两代硬件间 2.86 倍的差距说明 Blackwell Ultra 在 FP4 精度和 NVL72 互联上的收益是实打实的。
榜单同时印证了一点:软件优化能在已有硬件上继续榨出速度,并非只有换新卡才能提速。对采购方来说,GB300 与 H100 的单位训练成本对比如今有了可直接引用的数据。
MLPerf Training v6.1 results are out!
On 64 @NVIDIAAI Blackwell Ultra GPUs in a single NVIDIA GB300 NVL72 system, GPT-OSS-20B trained in 19.697 minutes. That is 2.86× faster than 64 NVIDIA H100 GPUs last round (Cisco) and 1.37× faster than 64 NVIDIA Blackwell GPUs (CoreWeave).
We also ran our first multimodal benchmark. In an unofficial reproduction, Flux.1 trained in 26.36 minutes on the same system, 1.50× faster than the last round's fastest on that hardware.
New hardware shortens training, and better software keeps finding speed on hardware you already have.
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include MLPerf, NVIDIA, GB300, Blackwell Ultra.
Source attribution: curated from Lambda (@LambdaAPI).