Google TPUs set new AI training performance benchmarks
In the latest round of MLPerf Training v6.1 benchmarks, Google Cloud presented 10 results that showcase the capabilities of its seventh-generation Ironwood TPUs and NVIDIA Grace Blackwell GPUs. These benchmarks spanned both frontier pre-training and marked the debut of an MLPerf benchmark for agentic reinforcement learning (RL). The results were achieved using Google Cloud’s AI Hypercomputer architecture, highlighting the collaborative design efforts across silicon, networking, and open software.
These benchmarks provide insight into the performance improvements in AI training on Google’s cloud infrastructure. The integration of advanced technology across hardware and software platforms is central to these achievements.
