Lifestyle

Google Cloud Publishes MLPerf Training v6.1 Results: 8,192 Ironwood TPUs Hit Target in 1.70 Minutes, Plus the First Agentic Reinforcement Learning Test

Google Cloud says it submitted 10 results to the MLPerf Training v6.1 AI benchmark using Ironwood TPUs and NVIDIA Grace Blackwell GPUs, covering pretraining and the first agentic reinforcement learning benchmark. This is a results announcement, not a product launch; all figures are Google's own.

News date: About 4 min read

Share on LINE
Contents
  1. 1. What results did Google publish?
  2. 2. How do the submissions map to scale and result IDs?
  3. 3. What do these numbers mean for general readers and developers?

In a blog post dated November 16, 2026, Google Cloud published 10 results from MLPerf Training v6.1. MLPerf is a benchmark that measures how fast AI models train, and the name is a trademark of the MLCommons association. The results cover Ironwood TPUs (Google's chips for AI computing) and NVIDIA Grace Blackwell GPUs (NVIDIA's processors used for AI work). They also include the first MLPerf agentic reinforcement learning (RL) benchmark. The post announces benchmark results, not a product launch or a new plan, and it does not state availability or pricing. Businesses and developers can treat the numbers as a performance reference for large-scale AI training, but all figures are Google's own.

What results did Google publish?

Google says the 10 results are built on Google Cloud's AI Hypercomputer architecture, reflecting co-design of silicon, networking and open software. The results fall into three groups: Ironwood pretraining (the stage where a model builds its base abilities from large amounts of data), Ironwood agentic RL (a training method in which the model produces outputs and then improves based on them), and NVIDIA Grace Blackwell GPUs.

  • Pretraining: Google says 8,192 Ironwood TPUs reached the benchmark target quality for the target model in 1.70 minutes, with 96% scaling efficiency at 4,096 chips.
  • Agentic RL: Google says 640 Ironwood TPUs reached the coding benchmark target accuracy in 18.02 minutes (0.765, target 0.69). Code generation and learning run at the same time, and training chips wait less than 3% of the time.
  • NVIDIA GPUs: Google says 256 GB300 GPUs on A4X Max instances (Google Cloud machines that customers rent) reached target quality in 31.96 minutes. There were also five more agentic RL submissions on A4X and A4X Max.

How do the submissions map to scale and result IDs?

Closed-division result IDs compiled from the footnotes of Google's post
Result IDHardware and scaleBenchmark
6.1-0017Ironwood, 640 chipsQwen3.5-397B RL
6.1-0018Ironwood, 2,048 chipsDeepSeek-V3
6.1-0019Ironwood, 4,096 chipsDeepSeek-V3
6.1-0020Ironwood, 8,192 chipsDeepSeek-V3
6.1-0021 / 0023 / 0025A4X, 56 / 64 / 256 GB200 GPUsRL
6.1-0022 / 0024A4X Max, 56 / 64 GB300 GPUsRL
6.1-0026A4X Max, 256 GB300 GPUsDeepSeek-V3

What do these numbers mean for general readers and developers?

This kind of benchmark shows how long it takes to train a large model to a specified quality under fixed rules. It is a useful reference for comparing hardware scale, but it does not mean your own model or data will get the same results. Google's post offers no comparison with other vendors or earlier products, and it does not describe actual rental terms. If you are evaluating adoption, you still need to confirm pricing, regions and availability separately.

Frequently asked questions

Is this a new product launch?

No. Google published MLPerf Training v6.1 benchmark results. The post does not mention new plans, pricing or availability regions.

What is special about the agentic reinforcement learning benchmark?

Google says it is the first MLPerf agentic reinforcement learning benchmark. Google says that when running on Ironwood, code generation and learning happen at the same time.

Where can I check the result IDs?

The footnotes of Google's post list Closed-division result IDs 6.1-0017 to 6.1-0026 and point readers to mlcommons.org to learn more.

Browse the latest news in this topic

Found this guide helpful? Share it with someone who needs it

Share on LINE

Sources

Compiled from the references and verification dates listed above; fares, schedules, product features, and terms can change, so please confirm the latest details on the relevant official website before travelling, booking, or purchasing.

Lifestyle