The MLPerf Training v6.1 results are out. The industry consortium MLCommons published the new round of the most important cross-vendor benchmark for AI training on 16 November 2026. For the first time it includes Nvidia’s Rubin generation and Google’s seventh-generation TPU, Ironwood. Also new is a test for post-training language models with reinforcement learning, designed to reflect how AI agents are built today.
MLPerf Training v6.1: record participation and two new tests
According to MLCommons, 24 organisations submitted a total of 153 systems, producing 250 test results. 57 systems, or 37 percent, were cloud-based, up from 29 percent in the previous v6.0 round. Twelve systems used 1,000 or more accelerators, the largest 10,240. Chips from Nvidia, AMD and Google were used. Cloud provider Crusoe took part as a submitter for the first time.
Two benchmarks are new:
- DLRMv4-HSTU replaces the previous DLRM-dcnv2 recommender test. It uses the attention-based HSTU architecture and the Yambda dataset, one of the largest open recommendation datasets.
- Qwen3.5-397B is, according to MLCommons, the first test of large-scale post-training using Group Relative Policy Optimization (GRPO). The model learns to solve software engineering tasks, including through tool calls.
The second item is the genuinely interesting one. Until now, MLPerf mainly measured how quickly a model is trained from scratch. Many advances of recent months, such as agentic capabilities, come from post-training instead. There, text generation, scoring and weight updates alternate constantly, which stresses systems differently from classic pre-training.
Nvidia: Vera Rubin doubles performance per GPU
Nvidia submitted Vera Rubin NVL72 for the first time. In Nvidia’s technical blog, the company cites the following closed-division figures:
- DeepSeek-V3 671B: 4.14 minutes on 1,024 GPUs across 16 racks. GB300 NVL72 needed 8.03 minutes in the same configuration, close to twice the performance per GPU.
- Qwen3.5-397B (GRPO): 10.44 minutes on 1,024 GPUs, compared with 13.78 minutes for GB300 NVL72. With the alternative vLLM inference backend it was 10.94 minutes.
- Other benchmarks: Vera Rubin is said to be 1.4 to 1.6 times faster per GPU than GB300 NVL72.
- Scaling: above 97 percent efficiency when going from 256 to 1,024 GPUs on DeepSeek-V3.
According to Nvidia, Vera Rubin delivered the fastest time to train on six of seven benchmarks. On the seventh, DeepSeek-V3, the absolute fastest result of the round also came from Nvidia, but from a GB300 cluster with 10,240 GPUs: 1.38 minutes. Around 80 percent of all closed-division submissions ran on Nvidia platforms, the company says. Nvidia also published figures from reruns with newer software. These are explicitly not verified by MLCommons and should be read with corresponding caution.
Why the RL test also stresses the CPU
The new GRPO benchmark shows how requirements for training systems are shifting. In reinforcement learning for agents, the model first produces solution attempts, so-called rollouts. These are executed in simulated environments, for example by running code or validating API calls, and then scored. Only then are the weights updated. Part of this work runs not on the GPU but on conventional processors.
Nvidia therefore highlights the role of its Vera CPU. A single Vera CPU rack with 256 CPUs ran more than 22,500 RL environments concurrently, the company says. Whether this figure is representative of other workloads cannot be judged from the blog. But it makes clear that agentic training no longer depends only on raw accelerator compute, but on the interplay of CPU, GPU and network.
Google: Ironwood in pre-training and agent training
Google submitted ten results on Ironwood TPUs and Nvidia GPUs in Google Cloud. According to the Google blog, 8,192 Ironwood TPUs reached the quality target in the pre-training benchmark in 1.70 minutes, with 96 percent scaling efficiency up to 4,096 chips. In the new RL benchmark, 640 Ironwood TPUs reached the target in 18.02 minutes. Code generation and learning ran simultaneously, so the training chips waited less than 3 percent of the time, Google says.
The figures do not contradict Nvidia; they simply measure different things. Nvidia’s 1.38 minutes on 10,240 GPUs and Google’s 1.70 minutes on 8,192 TPUs are different system sizes. So are Nvidia’s 10.44 minutes in the RL test on 1,024 GPUs and Google’s 18.02 minutes on 640 TPUs. Each vendor highlights the metric in which it leads: Nvidia per-GPU performance and the absolute best time, Google the efficiency of its TPUs.
What the results do and do not tell you
MLPerf is valuable because the tests are open, peer-reviewed and identical for all participants. Still, a sober look is worthwhile. The results show how quickly a system reaches a defined quality target. They say nothing about price, availability or power consumption per training run unless these are measured separately. Anyone comparing systems should filter for similar scale. MLCommons offers new visualisation tools in its results database for exactly that.
It is also notable who is missing. Amazon is not on the list of submitters, so AWS’s own training chips are not represented. Microsoft takes part through Azure, according to Nvidia with a GB200 cluster of 10,240 GPUs. The benchmark therefore mainly reflects competition between Nvidia, Google and AMD, with AMD playing hardly any role in the vendor blogs for this round.
For Nvidia, the results come at a turbulent time. After the change at the top of the company and the new US AI chip remote access rule, the company now has a technical argument: in an independently organised comparison, the new generation is significantly faster than the old one. How much of that reaches everyday large training runs depends on software, networking and utilisation. Data centre operators will still read the numbers closely, because halved training times mean more experiments per quarter on the same floor space.
Conclusion
MLPerf Training v6.1 shows two trends. First, AI training is becoming more diverse: more cloud systems, more vendors and, for the first time, a test specifically for agentic post-training. Second, Nvidia remains the benchmark but has to reckon with Google, which keeps pace with Ironwood at very large scale. The vendor figures in the blogs are carefully selected: the official submissions are verified by MLCommons, the later reruns are not. Anyone using them for their own decisions should go straight to the MLCommons results tables.




