Nov 16, 2026
In the latest MLPerf Training v6.1 benchmark round 1 , Google Cloud published 10 results across seventh-generation Ironwood TPUs and NVIDIA Grace Blackwell GPUs, spanning frontier pre-training and the first MLPerf benchmark for agentic reinforcement learning (RL). Built on Google Cloud’s AI Hypercomputer architecture, these results reflect co-design across silicon, networking, and open software:
- Pre-training on Ironwood: 8,192 Ironwood TPUs reached the target model's benchmark target quality in 1.70 minutes, scaling up to 4,096 chips with 96% efficiency.
- Agentic RL on Ironwood: 640 Ironwood TPUs reached the target model’s coding benchmark target accuracy in 18.02 minutes (0.765 vs. the 0.69 goal), running code generation and learning simultaneously so training chips waited under 3% of the time.
- NVIDIA Grace Blackwell GPUs: On Google Cloud A4X Max Instances, 256 NVIDIA GB300 GPUs reached the target model's benchmark quality in 31.96 minutes, complemented by five A4X and A4X Max submissions on the agentic RL benchmark.
For more details on the MLPerf Training v6.1 results, click here.
1MLPerf Training v6.1 Closed-division result IDs: 6.1-0017 (Ironwood 640-chip Qwen3.5-397B RL), 6.1-0018 (Ironwood 2,048-chip DeepSeek-V3), 6.1-0019 (Ironwood 4,096-chip DeepSeek-V3), 6.1-0020 (Ironwood 8,192-chip DeepSeek-V3), 6.1-0021 (A4X 56-GPU GB200 RL), 6.1-0022 (A4X Max 56-GPU GB300 RL), 6.1-0023 (A4X 64-GPU GB200 RL), 6.1-0024 (A4X Max 64-GPU GB300 RL), 6.1-0025 (A4X 256-GPU GB200 RL), and 6.1-0026 (A4X Max 256-GPU GB300 DeepSeek-V3). All reported times measure MLPerf time to train from the official reference checkpoint to the target convergence criterion. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. See mlcommons.org for more information
Posted in: