Train bigger and serve faster on the only AI cloud leading MLPerf in both training and inference

MLPerf is an industry-standard benchmark suite developed by MLCommons to measure and compare machine learning training and inference performance across hardware and platforms.

MLPerf benchmarks provide a fair, transparent method for evaluating AI hardware and cloud platforms, helping businesses choose solutions that offer optimal performance, scalability, and cost efficiency.

MLPerf Training v6.0 measures how quickly a computing system can train complex machine learning models—such as Meta's Llama 3.1 405B and the new DeepSeek-V3 671B mixture-of-experts benchmark—from initialization to a specified quality target, enabling fair and transparent performance comparisons across hardware platforms and cloud providers.

MLPerf Inference v5.0 measures how quickly computing systems process inputs and generate outputs using fully trained machine learning models, focusing specifically on throughput (tokens per second) and latency across realistic deployment scenarios to evaluate and compare the inference performance of hardware and cloud infrastructure providers.

In MLPerf Training v6.0, CoreWeave set new records—training DeepSeek-V3 671B in just 2.02 minutes on 8,192 NVIDIA GB300 GPUs, the fastest time in the round, and improving its Llama 3.1 405B time-to-train 2.8x year over year. These results were achieved on the same production GB300 NVL72 infrastructure customers run on every day.

CoreWeave's results come from the full platform, not hardware alone. By combining the latest NVIDIA GB300 NVL72 systems with NVIDIA Spectrum-X networking, the topology-aware SUNK scheduler, and CoreWeave Mission Control, CoreWeave sustains industry-leading performance and scaling efficiency from 64 to 8,192 GPUs—on the same infrastructure available to customers.