Google's Ironwood TPUs Post MLPerf v6.1 Results: 8,192 Chips, 1.70 Minutes

A single server tower unit with a perforated metal front panel in a blue-lit room

Photo: Hippopx (public domain). Illustrative image.

What happened. Google Cloud published 10 MLPerf Training v6.1 results on 16 November 2026, spanning its 7th-generation Ironwood TPUs and NVIDIA Grace Blackwell GPUs. The submission covers frontier pre-training and, for the first time in MLPerf, an agentic reinforcement-learning benchmark — a test of the training pattern behind today’s coding agents rather than the classic pre-training run. All figures below are Google-reported.

The headline numbers. With 8,192 Ironwood TPUs, Google reached the DeepSeek-V3 pre-training target in 1.70 minutes, with scaling efficiency of 96% at 4,096 chips. On the new agentic reinforcement-learning benchmark, 640 Ironwood TPUs hit the Qwen3.5-397B coding target in 18.02 minutes, scoring 0.765 against a goal of 0.69 — with the training chips waiting less than 3% of the time. For comparison, 256 NVIDIA GB300 GPUs on Google’s A4X Max system reached the pre-training target in 31.96 minutes. These are different systems at different scales, not a head-to-head chip ranking, and Google presents them that way. MLCommons published 250 results from 24 organisations in this round, including the first results from Rubin’s generation of GPUs. For the other half of the hardware race, see NVIDIA’s Rubin platform, announced at SC26.

Why the agentic benchmark matters more than the speed. Pre-training speed is a mature contest; the new benchmark is not. Agentic reinforcement learning — training a model by letting it act, fail and retry on coding tasks — is the technique behind the coding agents that Microsoft’s new MAI Code 2 family and OpenAI’s GPT-6.2 Sol are built to serve. A standardised way to measure that training matters because it is where the compute is increasingly going: not one giant pre-training run, but many iterative agent-training loops. Google’s 18-minute result is the first public number on that yardstick, and the efficiency figure — chips idle under 3% of the time — is the part infrastructure buyers will actually study.

Why it matters. MLPerf rounds are the closest thing AI hardware has to a shared scoreboard, and this one shows the contest splitting in two: raw pre-training scale on one side, agentic training efficiency on the other. Google’s Ironwood numbers are strong on both, but the lasting value of this submission is the benchmark itself — the industry now has a standard way to measure the training that makes agents, not just the training that makes models.

Sources

  • Google Ironwood
  • MLPerf v6.1
  • TPU
  • AI benchmarks
  • agentic reinforcement learning

Related articles