Google Cloud has published 10 results in the latest MLPerf Training v6.1 benchmark round, spanning its seventh-generation Ironwood TPUs and NVIDIA Grace Blackwell GPUs, including the first MLPerf benchmark for agentic reinforcement learning.
On pre-training, 8,192 Ironwood TPUs reached the target model's benchmark quality in 1.70 minutes, scaling up to 4,096 chips with 96% efficiency. On the new agentic reinforcement learning benchmark, 640 Ironwood TPUs reached the coding benchmark's target accuracy in 18.02 minutes, scoring 0.765 against a 0.69 goal, with training chips waiting under 3% of the time while code generation and learning ran together.
The NVIDIA side ran on Google Cloud's A4X Max instances, where 256 GB300 GPUs reached pre-training target quality in 31.96 minutes, alongside five further submissions on the agentic reinforcement learning benchmark.
The results are built on Google Cloud's AI Hypercomputer architecture, which the company describes as co-design across silicon, networking and open software.





