Google Cloud published ten benchmark results for its seventh-generation Ironwood TPU on November 17. The one worth reading twice is not the largest cluster — it is the first MLPerf benchmark built specifically for agentic reinforcement learning.
The numbers
A 640-chip Ironwood cluster reached target accuracy on a coding benchmark in 18.02 minutes, with chip idle time below 3%. Google also reported results across frontier pretraining workloads.
Idle time is the number practitioners should focus on. In large training runs, the difference between a good and a mediocre cluster is rarely peak throughput — it is how much of the time the expensive chips are waiting for data, for a collective operation, or for a straggler. Under 3% is a claim about the interconnect and the scheduler as much as the chip.
Why agentic RL is a different benchmark
Reinforcement learning for agents has an awkward shape. Instead of one enormous matrix multiplication repeated for weeks, it runs thousands of short rollouts: the model acts, an environment responds, a reward comes back, the weights update. The environment is often a container running a real program, which means the cluster spends time waiting on things that are not matrix math.
That is why the usual training benchmarks flatter hardware that is bad at this. A chip optimised for dense matmul can be excellent at pretraining and mediocre at agentic RL, because the bottleneck moves to scheduling and to keeping thousands of small jobs fed.
MLPerf adding this category matters more than any single score. It gives buyers a number that reflects the workload they are actually starting to run.
What Google is really selling
TPUs have always been sold on total cost of ownership rather than peak FLOPS, and this release continues that: a rack-scale number, an idle-time number, and a benchmark that matches where training budgets are moving.
The strategic context is the same as Nvidia's. Both companies are arguing that the constraint is now the cluster, not the chip, and both are publishing utilisation metrics because that is where the differentiation is.
The caveats
Benchmarks are configured by the vendor and run on the vendor's cluster. The 640-chip figure is a specific size that happens to be a clean multiple of a pod; results at 6,400 chips will not scale linearly, and nobody claims they will.
There is also the software question. JAX on TPU is excellent if your model fits the supported operations and painful if it does not. Teams with heavy custom kernels should test before committing, and the test should be their own model, not a benchmark suite.
The bigger picture
Three things are converging: training is shifting toward many short rollouts, inference is becoming the dominant cost, and power per rack is the binding constraint. A TPU generation tuned for agentic RL is a bet on all three at once.
If that bet is right, the interesting comparison next year will not be peak FLOPS but cost per completed agent episode — a metric nobody publishes yet, and the one that will decide which clusters get built.