HomeTech NewsData Centres & EnergyMLPerf Training v6.1 Adds First Rubin GPUs and TPUv7

MLPerf Training v6.1 Adds First Rubin GPUs and TPUv7

MLCommons has published results for MLPerf Training v6.1, the latest round of its peer-reviewed suite of full-system AI training tests. Participants submitted 153 systems, yielding 250 test results, from 24 organisations. Fifty-seven of those systems, or 37% of the total, were cloud-based, up from 29% in v6.0. Twelve systems used 1,000 or more accelerators, with the largest using 10,240. The systems used NVIDIA, AMD and Google accelerator hardware, MLCommons said, “including for the first time NVIDIA Rubin-generation GPUs and the Google TPUv7 (Ironwood).”

The release is dated 16 November 2026. It is a benchmark publication, not a product launch, and the official counts sit beside performance claims issued the same day by NVIDIA.

Two new tests, a larger field, and first-time hardware in the suite

MLPerf Training is an open-source, peer-reviewed suite of full system tests that, MLCommons says, “stress models, software, and hardware for a range of machine learning (ML) applications.” Version 6.1 adds two new benchmarks: DLRMv4-HSTU, a recommendation test that updates and replaces the earlier DLRM-dcnv2 benchmark and trains on the Yambda dataset; and Qwen3.5-397B, which MLCommons describes as “a first-of-its-kind workload for MLPerf Training: large-scale post-training using Group Relative Policy Optimization (GRPO).”

Shriya Rishab, a working-group co-chair, said the community had spent two years building the post-training test because “post-training has added a new level of complexity to the architecture of AI training.” Pavan Yalamanchili, the other co-chair, said the round had a higher level of participation and a wider mix of hardware, configurations and deployments than before.

The 24 submitting organisations listed by MLCommons are AMD, ASUSTeK, Azure, Cisco, CoreWeave, Crusoe, Dell, Firmus, Fujitsu, Giga Computing, Google, HPE, Inventec, KRAI, Lambda, MiTAC, Nebius, NVIDIA, Oracle, Quanta Cloud Technology, Supermicro, tiny corp, TTA and the University of Florida. David Kanter, head of MLPerf at MLCommons, particularly welcomed Crusoe as a first-time Training submitter. Full results and new visualisation tools are on the Training benchmark pages.

NVIDIA’s claims for Vera Rubin, and how they sit beside the official counts

NVIDIA’s blog, published the same day, says the company “was the only one with results on every benchmark this round,” covering large language models, image generation, recommender systems and reinforcement-learning post-training. In its debut, NVIDIA said, Vera Rubin NVL72 “delivered up to 2x more performance per GPU than the highly optimized NVIDIA GB300 NVL72 system.” Those are NVIDIA’s claims about its own submissions, not a ranking issued by MLCommons.

NVIDIA also said Vera Rubin NVL72 scaled on the DeepSeek-V3 benchmark “from 256 to 1,024 GPUs with about 97% efficiency: quadrupling the GPUs cut time to train from 16 minutes to about four.” It said it ran the same benchmark on 10,240 GPUs with GB300 systems, “representing the largest scale in this round,” and that Microsoft Azure “matched that scale with 10,240 NVIDIA GPUs with GB200 systems.” MLCommons confirms that the largest system in the round used 10,240 accelerators; the GB300 and GB200 identifications come from NVIDIA.

The first appearance of Rubin-generation GPUs in this suite is useful context for earlier TSN coverage of the same platform, including NVIDIA’s Rubin HPC science GPU, due in 2028, Oracle’s Vera Rubin launch on OCI, and NVIDIA’s earlier move to open Vera Rubin rack designs to OCP. Those posts are about products, racks and cloud availability. This one is about measured training runs in a common suite. A related, but separate, MLCommons exercise is the AILuminate safety ranking of 27 AI systems, which tests models for jailbreaks rather than training throughput.

A first appearance in a benchmark is not a TSN ranking

The firm facts from MLCommons are the size and shape of the round: 153 systems, 250 results, 24 organisations, a rise in the cloud share to 37%, twelve systems with at least 1,000 accelerators, a largest system of 10,240 accelerators, two new tests including a first post-training GRPO workload, and the first inclusion of NVIDIA Rubin-generation GPUs and Google’s TPUv7. Those are the official counts. They do not, by themselves, tell a reader which vendor “won,” or by how much.

NVIDIA’s “up to 2x” figure and the DeepSeek-V3 scaling result of “about 97% efficiency” are the company’s own account of its Vera Rubin NVL72 submissions, compared with its own GB300 NVL72 system. They should be read as vendor claims about a benchmark run, not as independent measurements published by TSN or as MLCommons commentary. The same applies to NVIDIA’s statement that it was the only submitter with results on every test in the round: it is a claim about coverage, and it is NVIDIA’s.

A first appearance of Rubin GPUs or TPUv7 in MLPerf Training is also not the same event as a hardware announcement. TSN has already covered Rubin HPC as a science GPU planned for 2028, Vera Rubin racks at OCP, and Oracle’s cloud launch. Those stories remain the place to go for product plans and availability. This publication says those accelerators have now been entered, for the first time, in this particular training suite, at a moment when more of the submitted systems are in the cloud and a handful have crossed a thousand accelerators.

The next official facts to watch are ordinary ones: whether later MLCommons rounds repeat the same hardware at larger scale, whether Google or AMD discuss their own v6.1 entries in a comparable way, and whether customers cite these results when they buy clusters. Until then, the news is that MLPerf Training v6.1 is out, that the field and the hardware mix have widened, and that NVIDIA has attached its own performance claims to the debut of Vera Rubin in the suite.

Sources

Related stories

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

More in this category

Latest stories

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.