The Pulse
Google reports Ironwood TPU results in MLPerf training tests
Google says an 8,192-chip Ironwood system reached a training benchmark target in 1.70 minutes. Its submissions also include an agentic reinforcement-learning test and results using Nvidia GPUs.

An 8,192-chip Ironwood system reached a model-training benchmark target in 1.70 minutes, Google said in a new set of MLPerf results. Google Cloud published 10 results on November 16 covering its seventh-generation TPUs and Nvidia Grace Blackwell GPUs.
The submissions span pre-training and agentic reinforcement learning, including what Google describes as the first MLPerf benchmark for agentic reinforcement learning. The tests use the company's AI Hypercomputer architecture.
Training at scale
The DeepSeek-V3 pre-training submissions used Ironwood configurations of 2,048, 4,096 and 8,192 chips. Google says the pre-training work scaled up to 4,096 chips with 96% efficiency.
The reported times measure training from an official reference checkpoint to the benchmark's target convergence criterion, rather than the entire process from the beginning of a model's training. The results are listed in MLPerf Training v6.1's closed division.
Agentic learning test
The Ironwood reinforcement-learning submission used 640 chips and the Qwen3.5-397B model. The company reports that the system reached the coding benchmark's target accuracy in 18.02 minutes, with a result of 0.765 against a goal of 0.69.
Google says code generation and learning ran simultaneously in that test, leaving the training chips waiting for under 3% of the time. Reinforcement learning can involve both sampling to generate new data and learner steps that update a model, according to an earlier Google Cloud description of the Ironwood software stack.
GPU submissions
On Google Cloud A4X Max instances, 256 Nvidia GB300 GPUs reached the DeepSeek-V3 benchmark quality target in 31.96 minutes, according to the company. Google also submitted five A4X and A4X Max results for the agentic reinforcement-learning benchmark.
Those tests included A4X configurations with 56 and 64 GB200 GPUs, and A4X Max configurations with 56 and 64 GB300 GPUs. A further A4X submission used 256 GB200 GPUs for reinforcement learning.
The Ironwood system
Ironwood is Google's seventh-generation Tensor Processing Unit. In an earlier technical guide, Google Cloud described an architecture that supports pods of up to 9,216 chips using inter-chip connections, optical switching, a data-center network and high-bandwidth memory.
Its optical circuit switch connects groups of chips, while inter-chip links connect chips within those groups. An earlier Google Cloud account says the optical network can bypass an unhealthy group or link and connect healthy groups instead.
Google's guide also describes software used for training on Ironwood, including JAX and MaxText. It identifies Tokamax as a library of JAX kernels optimized for TPUs and describes tools for distributing model weights, gradients and other training data across chips.
Sources
Companies in this story
Related stories
- NVIDIA reports Vera Rubin gains in MLPerf Training debut
- MLPerf Training adds agentic and recommender tests in v6.1
- Nvidia Sets a 51% Non-China Sales Threshold for Cloud Clients
- Google details bitwise-matched RL training across 64 TPU chips
- Google engineer joins shared stewardship of Agent Skills and MCP