NVIDIA says its Blackwell platform swept the results in MLPerf Training 6.0, claiming it led across every category in the benchmark suite.

In the same update, NVIDIA also highlighted its largest-scale training submission. The company said that run used 8,192 GPUs, positioning it as a top-end test of how the hardware handles big distributed training jobs.

There’s another point NVIDIA is leaning on: it says it was the only platform with submissions across all seven benchmarks in the MLPerf Training 6.0 suite. That matters because doing well on one or two tasks doesn’t necessarily mean a system is strong across a broader set of training workloads.

For people tracking how AI infrastructure is evolving, these kinds of benchmark rollups are one of the few ways to compare approaches under the same test umbrella. Still, it’s worth keeping expectations grounded: benchmark results describe performance on the specific set of tasks and rules defined by the test, not every real-world workload.

Even so, NVIDIA’s claim is straightforward. The company says Blackwell didn’t just rank near the top—it led “across every category” in MLPerf Training 6.0, with at least one submission scaling to thousands of GPUs and coverage spanning the full set of benchmarks in that training suite.

Source: NVIDIA Blog