RESEARCH · RESEARCH · #3679
Google Cloud publishes MLPerf Training v6.1 results for Ironwood TPUs and NVIDIA Grace Blackwell GPUs
Google Cloud published ten MLPerf Training v6.1 closed-division results running on its AI Hypercomputer architecture. An 8,192‑chip Ironwood TPU run reached the pre‑training target in 1.70 minutes (with reported scaling to 4,096 chips at 96% efficiency); a 640‑chip Ironwood run hit the agentic RL coding benchmark in 18.02 minutes (0.765 vs. 0.69 target) with training chips idle under 3% of the time. Separately, 256 NVIDIA GB300 (Grace Blackwell) GPUs on A4X Max instances reached the DeepSeek‑V3 target in 31.96 minutes, and multiple A4X/A4X Max submissions were reported for the new agentic RL benchmark.
KEY POINTS
- Google Cloud published ten MLPerf Training v6.1 closed-division results running on its AI Hypercomputer architecture.
- An 8,192‑chip Ironwood TPU run reached the pre‑training target in 1.70 minutes (with reported scaling to 4,096 chips at 96% efficiency); a 640‑chip Ironwood run hit the agentic RL coding benchmark in 18.02 minutes (0.765 vs.
- 0.69 target) with training chips idle under 3% of the time.
WHY IT MATTERS
These results demonstrate high-scale training speed and efficiency of Google’s Ironwood TPUs and provide the first MLPerf agentic RL benchmark data, informing choices for large-scale model training and infrastructure.