← The feed Many Minded · The brief
KNOWLEDGE · forward · impact 3/5 · 2026-11-17 · Google Cloud

Google posts Ironwood TPU results in latest MLPerf training round

Google Cloud's 10 MLPerf Training v6.1 entries cover Ironwood TPUs and NVIDIA Blackwell GPUs, including the first agentic reinforcement learning benchmark.

Google Cloud submitted 10 results to the MLPerf Training v6.1 round, dated November 16, 2026. They cover its seventh-generation Ironwood TPUs and NVIDIA Grace Blackwell GPUs, across frontier pre-training and the first MLPerf benchmark for agentic reinforcement learning.

On pre-training, a cluster of 8,192 Ironwood TPUs reached the target quality on the DeepSeek-V3 benchmark in 1.70 minutes. Google reports 96% scaling efficiency when growing to 4,096 Ironwood chips. On the new agentic RL benchmark, 640 Ironwood TPUs reached the coding target on Qwen3.5-397B in 18.02 minutes, scoring 0.765 against a 0.69 goal. Code generation and learning ran at the same time, and training chips sat idle less than 3% of the time. On A4X Max instances, 256 NVIDIA GB300 GPUs hit the DeepSeek-V3 pre-training target in 31.96 minutes. Five more A4X and A4X Max entries used 56, 64 and 256 GPU configurations on the RL benchmark.

The times measure MLPerf time to train, from the official reference checkpoint to the target convergence criterion. All entries are Closed-division results on Google Cloud's AI Hypercomputer architecture.

Why it matters: access to knowledge tools depends partly on how much compute it takes to train capable models. Faster time to train, high scaling efficiency and little idle hardware mean the same capacity can do more work, which over time pushes the cost of building and refining models down. Rentable cloud capacity also means smaller labs can reach this hardware without building it.

Caveats: these are Google Cloud's own reports of its entries, and nothing here compares them with other submitters. The benchmark times are not a direct measure of what it costs to train a frontier model, so the cost link is our inference. Watch for rival submissions in the same round and for any published pricing on Ironwood capacity.

Source: Google Keyword