NVIDIA says its Blackwell systems delivered the fastest training times across every benchmark in MLPerf Training 6.0. In the company’s telling, that means Blackwell was consistently the quickest option, not just strong in one or two test areas.
The update also points to the kind of scale people care about when they’re trying to run big training jobs. NVIDIA says its Blackwell NVL72 systems were used at a scale of 8,192 GPUs for the MLPerf Training 6.0 results.
NVIDIA further claims it is the only platform that submitted results across all seven MLPerf Training benchmarks. That matters because training performance is hard to compare when different vendors only publish partial test coverage. Having full coverage can make it easier to judge whether a platform’s advantage holds up across different training setups.
Why this is worth attention: for teams planning runs on large clusters, “fast” isn’t only about one workload. They want a system that can keep up across a range of training benchmarks, and they also need evidence that it can scale to thousands of GPUs without performance turning into a moving target.
That said, it’s still important to treat benchmark results as one data point, not a guarantee of future outcomes for every model, batch size, or network setup. Even when the test suite is the same, performance can shift with the details of the environment. Still, NVIDIA’s claim gives a clear picture of what Blackwell performed like under MLPerf Training 6.0’s rules.
Source: NVIDIA

