• The consortium achieved unprecedented benchmark performance in AI training and inference workloads.
  • Results demonstrate the GB200's capabilities in large language model processing and generative AI tasks.
  • The submission highlights growing industry adoption of NVIDIA's latest architecture.

A New Benchmark in AI Performance

CoreWeave, NVIDIA and IBM have collectively submitted what industry observers are calling the most impressive MLPerf benchmark results to date, showcasing the raw power of NVIDIA's GB200 Grace Blackwell Superchips. The tests, conducted across multiple configurations, demonstrate significant improvements in both training and inference workloads compared to previous generations.

Early analysis suggests the GB200-based systems completed certain large language model training tasks in nearly half the time of comparable H100-based clusters. One particularly striking result showed a 72-GPU GB200 NVL72 system delivering real-time inference for trillion-parameter models at speeds previously thought unattainable in production environments.

Technical Breakthroughs on Display

The benchmark submission highlights several architectural advantages of the Grace Blackwell platform. Its multi-chip module design, combining one Grace CPU with two Blackwell GPUs via NVLink-C2C interconnect, proved particularly effective in distributed training scenarios. The system's massive memory bandwidth - up to 16 TB/s for HBM3e GPU memory - allowed for efficient processing of enormous parameter sets without constant data shuffling.

"What we're seeing here represents more than incremental improvement," said one engineer familiar with the tests, speaking on condition of anonymity because the full results haven't been officially published. "The GB200's ability to treat 72 GPUs as a single domain changes the economics of large-scale AI deployment."

Industry Implications

While MLPerf results represent controlled benchmarks rather than real-world performance, the strong showing from CoreWeave, NVIDIA and IBM suggests the GB200 platform may accelerate enterprise adoption of generative AI technologies. Several major cloud providers are reportedly fast-tracking GB200-based offerings following the benchmark revelations.

The submission comes as competition intensifies in the AI accelerator market, with multiple vendors preparing next-generation offerings. However, these results demonstrate NVIDIA's continued leadership in raw performance for cutting-edge AI workloads, particularly in large language model applications.

Representatives from CoreWeave and IBM declined to comment ahead of the official MLPerf results publication, while NVIDIA pointed to its earlier statements about the GB200's architectural advantages. Full details are expected to be released in the coming weeks as MLPerf completes its verification process.