Newshot

Google's Seventh-Gen Ironwood TPU Delivers: 8,192-Chip Cluster Reaches DeepSeek-V3 Pretraining Target in About 102 Seconds

AI summary

In the first round of MLPerf Training v6.1, Google Cloud submitted 10 results spanning its seventh-gen Ironwood TPUs and 英伟达 Grace Blackwell GPUs, covering model pretraining and MLPerf's first agentic reinforcement learning test. In the DeepSeek-V3 pretraining test, 8,192 Ironwood TPUs took 1.70 minutes (about 102 seconds) to bring the target model to the benchmark's required quality level, and Google also disclosed a 96% scaling efficiency in a test scaled to 4,096 chips.

Translation(model translation, not the source's own text)

Article title: Google's 7th-gen in-house Ironwood TPU delivers: 8,192-chip cluster hits DeepSeek-V3 pretraining target in about 102 seconds

Body:

  • Day/Night Follow system Light Dark

Google's 7th-gen in-house Ironwood TPU delivers: 8,192-chip cluster hits DeepSeek-V3 pretraining target in about 102 seconds

IT Home, November 17 news: Google Cloud published a blog post yesterday (November 16), submitting 10 results in the first round of the MLPerf Training v6.1 benchmark, covering the 7th-gen Ironwood TPU and 英伟达 Grace Blackwell GPUs, involving model pretraining and 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 in-house chips, and the other 6 use 英伟达 GPUs.

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

According to the participant supplemental notes published by MLCommons, Google submitted three sets of Ironwood results in the DeepSeek-V3 671B pretraining project:

Number of Ironwood chipsTime to reach target qualityConverted to secondsSpeedup relative to the 2,048-chip configuration
2,048 chips5.03 minutesabout 302 secondsbaseline
4,096 chips2.62 minutesabout 157 secondsabout 1.92x
8,192 chips1.70 minutes102 secondsabout 2.96x

In the same round of DeepSeek-V3 671B tests, the following public results can be referenced:

Submitter/platformAccelerator configurationTime to target
英伟达 GB300 NVL72 cluster10,240 GB300 GPUs1.38 minutes, about 82.8 seconds
Google Ironwood cluster8,192 Ironwood TPUs1.70 minutes, 102 seconds
英伟达 Vera Rubin NVL72 cluster1,024 Rubin GPUs4.14 minutes
英伟达 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 英伟达's technical notes for this round, and 英伟达 lists the 1.38 minutes of 10,240 GB300s 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 running Qwen3.5-397B took 18.02 minutes to meet the programming task requirements, with an accuracy of 76.5%, above the 69% target. Code generation and learning proceed synchronously, keeping the share of training chip waiting time below 3%.

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

  • New AI training performance benchmarks for Google TPUs
  • MLCommons Releases MLPerf Training v6.1 Results, Spotlighting Recent Advances in AI Training
  • Supplemental Results Discussion for MLPerf Training v6.1

Advertising statement: External redirect links contained in the article (including but not limited to hyperlinks, QR codes, passwords, 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.

Related articles

  • Google Pics upgrades to the Nano Banana 2.1 model: more refined image generation, more accurate image editing
  • The best AI assistant for an app side hustle: Google launches a series of tools targeting the most troublesome monetization step
  • Addressing AI vulnerability discovery challenges: Google Project Zero plans to gradually shorten the public disclosure deadline for vulnerabilities to 30 days
  • Google Pixel series phones have problems after the November update: screensaver cannot enter low-brightness mode
  • Google pushes Android Auto 18.2 beta, code shows the official team is continuing to develop a light mode menu interface
  • Google Gemini personal intelligence feature connects to the Google Wallet app: can provide personalized analysis and suggestions for users' spending habits

Websites under Ruanmei: IT Home Zuihuimai - rebate and cashback coupons iPhone Home Win7 Home Win10 Home Win11 Home

Software under Ruanmei: Ruanmei Mobile APP Application Cube Zuihuimai Yaozhi

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗