Google Cloud submits 10 MLPerf results, Ironwood TPUs compete with Nvidia GPUs
Google Cloud submitted 10 results in the first round of MLPerf Training v6.1, covering its seventh-gen Ironwood TPUs and Nvidia Grace Blackwell GPUs. The tests include model pre-training and agentic reinforcement learning, which was added for the first time. Four results used Google's own chips, while six used Nvidia GPUs.
In the DeepSeek-V3 pre-training test, 8,192 Ironwood TPUs reached the target quality in 102 seconds. Scaling to 4,096 chips showed 96% system efficiency. Nvidia's GB300 NVL72 cluster with 10,240 GPUs finished in 82.8 seconds, the fastest time in this round for the project.
In the agentic reinforcement learning test, 640 Ironwood chips ran Qwen3.5-397B and met the coding task requirement in 18.02 minutes with 76.5% accuracy, above the 69% target.
Comments
Loading comments…