Google Ironwood TPU Cluster Trains DeepSeek-V3 in 102 Seconds

Original title: 谷歌第七代自研 Ironwood TPU 交卷:8192 颗集群约 102 秒达标 DeepSeek-V3 预训练
78AI Score

Google Cloud submitted 10 MLPerf Training v6.1 results, four of them on its seventh-generation Ironwood TPU. On DeepSeek-V3 671B pretraining, 8,192 Ironwood chips reached the target quality in about 102 seconds with 96% scaling efficiency at 4,096 chips, though Nvidia's 10,240-GPU GB300 cluster posted the fastest time at 82.8 seconds. The round also debuted an agentic reinforcement learning benchmark, where 640 Ironwood chips ran Qwen3.5-397B to 76.5% accuracy on a coding task.

Full Report

IT Home November 17 news, Google Cloud yesterday (November 16) published a blog post, submitting 10 results covering the seventh-generation Ironwood TPU and NVIDIA Grace Blackwell GPU in the first round of the MLPerf Training v6.1 benchmark, involving model pre-training as well as the agentic reinforcement learning test included in MLPerf for the first time.

IT Home note: The 10 results do not all come from Ironwood; 4 of them use Google's self-developed chips, while the other 6 use NVIDIA GPUs.

In the DeepSeek-V3 pre-training test, 8,192 Ironwood TPUs took 1.70 minutes (about 102 seconds) to bring the target model up to the quality requirements specified by the benchmark. Google also disclosed that in a test scaled to 4,096 chips, the system achieved 96% scaling efficiency.

According to the supplementary notes from participants published by MLCommons, Google submitted three sets of Ironwood results in the DeepSeek-V3 671B pre-training project:

Number of Ironwood chipsTime to reach target qualityConverted to secondsSpeedup vs. 2,048-chip configuration
2,048 chips5.03 minutes~302 secondsBaseline
4,096 chips2.62 minutes~157 seconds~1.92x
8,192 chips1.70 minutes102 seconds~2.96x

In the same round of DeepSeek-V3 671B testing, the following publicly available results can be referenced:

Submitter / PlatformAccelerator ConfigurationTime to Completion
NVIDIA GB300 NVL72 Cluster10,240 GB300 GPUs1.38 minutes, approx. 82.8 seconds
Google Ironwood Cluster8,192 Ironwood TPUs1.70 minutes, 102 seconds
NVIDIA Vera Rubin NVL72 Cluster1,024 Rubin GPUs4.14 minutes
NVIDIA GB300 NVL72 Cluster1,024 GB300 GPUs8.03 minutes
Google Cloud A4X Max256 GB300 GPUs31.96 minutes

The Ironwood and Google Cloud GB300 results come from Google's disclosure; the remaining results come from NVIDIA's technical briefing for this round, in which NVIDIA lists 1.38 minutes for 10,240 GB300 chips as the shortest time for this project in this round.

In the agentic reinforcement learning test introduced for the first time in this round, 640 Ironwood chips ran Qwen3.5-397B and reached the programming task requirements in 18.02 minutes, with an accuracy of 76.5%, above the 69% target. Code generation and learning proceed simultaneously, keeping the training chip wait time share below 3%.

In the Google Cloud A4X Max instance, 256 NVIDIA GB300 GPUs used for the pretraining benchmark reached the target quality in 31.96 minutes, supplemented by 5 submissions of A4X and A4X Max in the agentic RL benchmark test.

References

Advertising statement: External redirect links contained in the article (including but not limited to hyperlinks, QR codes, passcodes, and other forms) are used to convey more information and save selection time; the results are for reference only. All articles on IT Home that contain external links include this statement.