The results of 'MLPerf Training v6.0,' a benchmark test evaluating the learning performance of AI servers from NVIDIA and AMD, have been released.
MLCommons, an industry group that collects and publishes performance measurement results for AI infrastructure, released the results of its learning performance evaluation test, ' MLPerf Training v6.0, ' on June 16, 2026.
MLPerf Training v6.0 Results: New MoE Benchmarks and Record System Diversity | MLCommons
NVIDIA Blackwell Sweeps MLPerf Training 6.0 | NVIDIA Blog
https://blogs.nvidia.com/blog/blackwell-mlperf-training-6-0/
Technical Dive into AMD's MLPerf Training v6.0 Submission — ROCm Blogs
https://rocm.blogs.amd.com/artificial-intelligence/mlperf-training-v6.0/README.html
MLPerf Training is a benchmark test that measures the performance of AI models when training them on AI infrastructure. Chip manufacturers such as NVIDIA and AMD submit their measurement results to MLCommons, and third-party vendors such as Dell, Fujitsu, and Cisco also submit their results.
The MLPerf Training test content is updated to keep up with current trends, and in MLPerf Training v6.0, training with 'DeepSeek V3' and 'gpt-oss-20b' was added as a test item.
As of the announcement of the results on June 16, 2026, test results using a total of 95 different systems from 24 organizations have been submitted, and the results can be viewed at the following link. Overall, systems equipped with NVIDIA GPUs recorded high scores, with systems equipped with AMD GPUs following closely behind.
MLCommons MLPerf Training Benchmark
The results of MLPerf Training v6.0 are also being promoted on each company's official website. NVIDIA has published a graph comparing the training speed when training the same model with GB200 NVL72 and GB300 NVL72, and is promoting that 'GB300 NVL72 can train up to 1.6 times faster than GB200 NVL72.'
Microsoft also
AMD is promoting the fact that they can improve training speed by 3.5 times by upgrading from an environment using 'MI300X with MXFP8 precision' in 2025 to an environment using 'MI355X with MXFP4 precision' in 2026.
Furthermore, comparing the environment of 'training with NVFP4 precision on the B200' and the environment of 'training with MXFP4 precision on the MI355X,' they stated that 'the performance difference is within 5% for fine-tuning of Llama 2 70B and 6% for pre-training of Llama 3.1 8B,' emphasizing that they are maintaining their competitiveness with NVIDIA.
Related Posts:
