THE ESSENTIALS IN 30 SECONDS

  • MLCommons released MLPerf Training v6.1 on 16 November: 250 results from 24 organizations, including first results for NVIDIA Rubin-generation GPUs and Google's TPUv7 Ironwood.
  • The round adds a Qwen3.5-397B reinforcement-learning post-training benchmark (GRPO) and replaces the old DLRM recommender test with DLRMv4-HSTU.
  • NVIDIA says Vera Rubin NVL72 was fastest on six of seven benchmarks and up to 2x faster per GPU than GB300; AMD says MI355X matched or beat B200 in its comparisons. These ratios are vendor calculations.

MLCommons, the industry consortium behind the MLPerf benchmarks, published the results of MLPerf Training v6.1 on 16 November. The round drew 250 results from 24 organizations, among them AMD, Azure, Cisco, CoreWeave, Dell, Google, HPE, Lambda, Nebius, NVIDIA, Oracle, Supermicro, tiny corp and the University of Florida. Crusoe took part for the first time. According to MLCommons, the largest system used 10,240 accelerators, and 37% of systems were cloud-based, up from 29% in the previous round.

What changed in the benchmark

The suite now has seven tests. The main addition is a post-training workload in which Qwen3.5-397B is trained with Group Relative Policy Optimization, a reinforcement-learning method, on software engineering tasks involving tool calls. MLCommons also replaced its older recommender benchmark with DLRMv4-HSTU, built on an attention-based architecture. TechArena notes that the post-training test exercises the whole system, including the conventional processors that run sandboxed checks, rather than the accelerators alone.

The round includes the first results for NVIDIA's Rubin-generation GPUs and Google's TPUv7 Ironwood. One of the largest DeepSeek-V3 runs used 8,192 Ironwood chips.

What the vendors claim

NVIDIA says its Vera Rubin NVL72 rack posted the fastest time to train on six of the seven benchmarks. On DeepSeek-V3, the fastest absolute time came from a 10,240-GPU GB300 run. Comparing equal GPU counts, the company says Vera Rubin is between 1.4x and 2x faster per GPU than GB300, with the largest gap on DeepSeek-V3. NVIDIA also published reruns on newer software showing up to 2.4x, but says MLCommons has not verified them.

AMD submitted MI350X and MI355X systems on five benchmarks, up from three, including its first mixture-of-experts entry, GPT-OSS-20B. The company describes a FLUX.1 run using FP6 and FP4 formats as the first public, reproducible FP6 training demonstration. It says MI355X finished faster than NVIDIA's B200 in several eight-GPU tests while trailing B300. Partner submissions from Crusoe, Dell, HPE, MiTAC, Oracle and Supermicro used its chips.

How to read the numbers

MLCommons publishes raw times to train, not rankings. The per-GPU ratios and percentage comparisons above are calculations by NVIDIA and AMD, made on hardware, software and system sizes that differ between submissions. AMD states that its aggregate figures are not official MLPerf scores. The full results can be browsed on the MLCommons results site.