NVIDIA has emerged as the top performer in the latest MLPerf Training v5.1 benchmarks. These are industry recognized tests that measure how quickly and efficiently computer systems can train complex artificial intelligence models. NVIDIA's systems achieved the fastest times across all seven categories. This included large language models, image generation, recommender systems, computer vision, and graph neural networks. NVIDIA was the only company to submit results for every test, highlighting the flexibility of its GPUs and the power of its CUDA software tools.
A major factor in NVIDIA's success was the introduction of its new Blackwell Ultra GPUs, which are now used in systems like the GB300 NVL72. These GPUs brought huge improvements compared to the earlier Hopper GPUs, offering over four times the performance for some large language model training tasks. The Blackwell Ultra GPUs use advanced architectural designs, like new Tensor Cores and large amounts of high speed memory, to make training much faster.
Another breakthrough came from using NVFP4 precision, a way for the GPUs to perform calculations using fewer data bits. This method allows for quicker computations without sacrificing the quality of the results. NVIDIA is currently the only platform to submit MLPerf Training results using this low precision approach while still meeting accuracy standards.
NVIDIA set new records in training times. For example, a team of over five thousand Blackwell GPUs trained a huge language model in just ten minutes. This was nearly three times faster than previous records. NVIDIA also demonstrated that even with fewer GPUs, training times have improved significantly compared to past benchmarks.
Newer tests, like those for the Llama 3.1 8B and FLUX.1 models, were also included in this round. NVIDIA set the standard in these new categories as well, achieving very fast training times with large groups of Blackwell and Blackwell Ultra GPUs. The company continues to hold records in existing tests for graph neural networks, object detection, and recommender systems.
NVIDIA did not achieve this alone. Their technology partners participated by submitting strong results, showing the breadth of NVIDIA’s ecosystem. Organizations like Dell, Hewlett Packard Enterprise, Supermicro, and many others contributed to this round’s tests.
NVIDIA continues to drive rapid improvements in AI training performance each year, making AI development faster and more accessible. The company’s ongoing innovation helps pave the way for future breakthroughs in intelligent technology.
Original article and image: https://blogs.nvidia.com/blog/mlperf-training-benchmark-blackwell-ultra/
