NVIDIA published an update on its Blackwell platform on June 16, 2026, tying the chip’s performance to the MLPerf Training 6.0 benchmark results.

In the post, the company said Blackwell led across every category in the benchmark. NVIDIA also claimed it delivered the fastest time to train on every benchmark included in the suite.

Another point NVIDIA highlighted was scale. The company said it scaled one submission up to 8,192 GPUs using GB200 NVL72 systems. NVIDIA called this the largest Blackwell-based submission in the suite to date, which matters because training runs at this size are often where efficiency and system-wide performance show up most clearly.

These benchmark results are important mainly for what they suggest about how well a hardware platform can keep up as workloads get bigger, not just whether a single model trains quickly. Training at large GPU counts usually involves more than raw chip speed; the full setup has to work smoothly as communication and workload distribution grow more demanding.

What NVIDIA claims in MLPerf Training 6.0

According to NVIDIA, Blackwell:

• Led in every MLPerf Training 6.0 category.
• Achieved the fastest time to train across all benchmarks in the suite.
• Ran a submission at 8,192 GPUs using GB200 NVL72 systems.

NVIDIA’s post frames the latest run as a step up in both speed and scale for its Blackwell training approach, based on what the MLPerf suite is designed to measure.

Source: NVIDIA Blog