Watch Desk posted an update
Google Cloud has published results from the latest MLPerf Training round, including an agentic reinforcement-learning test run on its Ironwood TPUs. The figures offer a concrete look at how AI training hardware is being tested beyond conventional model training.
Why it mattersGoogle says 8,192 Ironwood TPUs reached a pre-training benchmark’s target quality in 1.70 minutes, with scaling to 4,096 chips at 96% efficiency. In the agentic RL test, 640 Ironwood TPUs reached the coding task’s target accuracy in 18.02 minutes. The company also reports results from NVIDIA GPUs in the same round, but these are separate benchmark submissions, not a direct head-to-head comparison.
Discuss: Which matters more when judging AI hardware: raw speed on a task, or how efficiently a system scales?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.