Google Cloud Reveals Advanced AI Training Benchmarks with Seventh-Generation Ironwood TPUs

Image: Google · Source
Google Cloud showcases new MLPerf Training v6.1 results with seventh-generation Ironwood TPUs and NVIDIA Grace Blackwell GPUs, demonstrating leading AI training speeds and efficiency on frontier pre-training and agentic reinforcement learning benchmarks.
Google Cloud has published new AI training performance benchmarks as part of the MLPerf Training v6.1 round, highlighting substantial advances using its seventh-generation Ironwood Tensor Processing Units (TPUs) and NVIDIA Grace Blackwell GPUs. These results showcase Google’s continued investment in co-designed silicon, networking, and open software infrastructure within its AI Hypercomputer architecture.
Among the key achievements, 8,192 Ironwood TPUs completed pre-training on a major machine learning model in just 1.70 minutes, scaling efficiently up to 4,096 chips with a sustained 96% efficiency. This rapid training time marks a significant milestone in reducing AI development cycles for large-scale models.
Additionally, the Ironwood architecture demonstrated its capabilities on a new agentic reinforcement learning (RL) benchmark—the first MLPerf contest of its kind. Using 640 Ironwood TPUs, the system reached the agentic RL target coding accuracy in 18.02 minutes, surpassing the benchmark goal score of 0.69 with a result of 0.765. The training setup also optimized chip utilization, ensuring hardware was idle less than 3% of the time during simultaneous code generation and learning tasks.
Alongside TPUs, Google Cloud also submitted results using NVIDIA Grace Blackwell GPUs on its A4X Max Instances. A cluster of 256 NVIDIA GB300 GPUs completed the pre-training benchmark in 31.96 minutes, with additional submissions targeting the agentic RL benchmark.
These benchmarks reflect the synergy between Google’s custom hardware and software stack, pushing the boundaries of AI training speed and efficiency. The detailed results underscore the benefits of tightly integrated cloud AI infrastructure tailored to large and complex model workloads.
For the full technical details and further outcomes from the MLPerf Training v6.1 round, Google Cloud’s official blog provides comprehensive coverage.
Sources and original reporting
Read the original source ↗

Comments (0)
No comments yet. Start the discussion.
Write a comment
Comments are published after moderation. Your name and comment will be visible publicly. Account