MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
MLCommons has published the results of MLPerf Training v6.1, the shared benchmark where vendors and clouds compare how fast they can train neural networks. A total of 153 systems submitted 250 results, and for the first time the list includes new Nvidia Rubin and Google TPU Ironwood accelerators. The suite also adds a fine-tuning test on Qwen3.5-397B using GRPO, the method used to train agents. Anyone choosing a cluster or renting GPUs gets an independent comparison under the same rules.
AI-processed from MLCommons; edited by Hamidun News
When a company decides what to train its model on, whether Nvidia, AMD or Google TPU, a single server or a cluster of thousands of accelerators, it is hard to trust vendor presentations: each one shows its own numbers on its own tasks. That is why MLPerf Training from the MLCommons consortium exists. It is a shared exam: all participants run the same tasks under the same rules, and their results are checked by their peers.
In the latest round, v6.1, 153 systems (that is, individual servers and clusters) submitted 250 results. For the first time the list includes new Nvidia accelerators of the Rubin generation and Google's seventh-generation TPU, called Ironwood.
A test for fine-tuning the Qwen3.5-397B model has also been added, covering the stage where AI agents are made today.
What changed among the participants
There were more participants than in the record-setting v6.0 round from half a year ago. Results came from 24 organizations, including AMD, Azure, Cisco, CoreWeave, Dell, Google, HPE, Lambda, Nebius, Nvidia, Oracle, Supermicro and others. The cloud provider Crusoe made its debut.
The cloud share is growing: 57 of the 153 systems, or 37%, run in the cloud. In v6.0 the figure was 29%. Twelve systems use 1,000 or more accelerators, and the largest has 10,240. By hardware, the list includes Nvidia, AMD and Google.
Working group co-chair Pavan Yalamanchili says the results now cover a much larger part of the "matrix of possible systems." That makes it easier to compare configurations than before.
Two new tests
The first is DLRMv4-HSTU, a test for recommendation systems, the part of AI that selects what goes into your feed, videos and products. It replaces the earlier DLRM-dcnv2. It is built on the HSTU architecture, which uses an attention mechanism, and is trained on the Yambda dataset, one of the largest open datasets for recommendations.
The second is Qwen3.5-397B, and for people building agents it is the more interesting one. It is the first post-training workload in MLPerf Training history: a large pretrained model is fine-tuned with GRPO (Group Relative Policy Optimization). This is how models are trained today to reason and call tools. In the test, the model is trained on software engineering tasks, including tool-calling. According to the organizers, the benchmark has been in development for the past two years, and the number of submissions shows that participants adopted it quickly.
What to do with this in practice
First, the industry now has a common way to compare hardware not only on classic training from scratch but also on fine-tuning. Fine-tuning is currently a major factor in agent quality. Its load profile is different: the system alternates between generating responses and updating the model's weights. So an accelerator that is strong at conventional training is not necessarily the best choice here. If you plan to fine-tune a large model for your own agents, this is the test to look at.
Second, the wide range of configurations helps find the right balance for a given budget, from a single node to a cluster of thousands of accelerators. The MLCommons website has an updated results page where you can compare systems and rank configurations of similar scale.
There are caveats too. The results are submitted by the vendors and clouds themselves, and they tune their systems specifically for the test. Your project may see different speeds. In addition, the MLCommons press release has no summary table of training times by system. Conclusions about the leaders will have to be drawn from the full data on the benchmark's website.
Still, for planning hardware purchases and GPU rentals, MLPerf remains one of the few independent reference points. The inclusion of Rubin and Ironwood gives the first opportunity to compare the new generation of accelerators under the same rules.
- Nov 21, 2026Nvidia doubled profit, but shares fell 2% over the search for a new CEO
- Nov 21, 2026Nous Research fine-tuned China's Qwen3.8, and propaganda in its answers dropped from 22% to 1%
- Nov 21, 2026Samsung Wants to Buy Modine, a Data Center Cooling Supplier, for Nearly $6 Billion
- Nov 21, 2026Reuters: China's $3.2B AI Data Center Buildout Relies on Nvidia Chips as Huang Steps Down
Key Questions Answered
What changed among the participants?
There were more participants than in the record-setting v6.0 round from half a year ago. Results came from 24 organizations, including AMD, Azure, Cisco, CoreWeave, Dell, Google, HPE, Lambda, Nebius, Nvidia, Oracle, Supermicro and others. The cloud provider Crusoe made its debut. The cloud share is growing: 57 of the 153 systems, or 37%, run in the cloud. In v6.0 the figure was 29%. Twelve systems use 1,000 or more accelerators, and the largest has 10,240. By hardware, the list includes Nvidia, AMD and Google.
What are the key details on Two new tests?
The first is DLRMv4-HSTU, a test for recommendation systems, the part of AI that selects what goes into your feed, videos and products. It replaces the earlier DLRM-dcnv2. It is built on the HSTU architecture, which uses an attention mechanism, and is trained on the Yambda dataset, one of the largest open datasets for recommendations. The second is Qwen3.5-397B, and for people building agents it is the more interesting one.
What to do with this in practice?
First, the industry now has a common way to compare hardware not only on classic training from scratch but also on fine-tuning. Fine-tuning is currently a major factor in agent quality. Its load profile is different: the system alternates between generating responses and updating the model's weights. So an accelerator that is strong at conventional training is not necessarily the best choice here. If you plan to fine-tune a large model for your own agents, this is the test to look at.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.