MLCommons published the MLPerf Training v6.1 results on November 16, and the round has two firsts: a benchmark for AI post-training with reinforcement learning, and the first results from Nvidia's Rubin-generation GPUs and Google's TPUv7 "Ironwood." MLCommons counts 24 submitting organizations and 250 results — its largest Training round so far.
Why it matters: MLPerf is the closest thing the AI-chip industry has to a common, peer-reviewed yardstick. Vendors' marketing slides compare their own chips on their own terms; MLPerf at least fixes the workloads and the accuracy targets. This round also changes what's being measured — from pretraining a model on a fixed dataset toward the reinforcement-learning loops labs now use to make models into agents.
The round in numbers
| Item | Figure (per MLCommons unless noted) |
|---|---|
| Submitting organizations | 24 (Crusoe new) |
| Results | 250 |
| Systems | 153 per the press release (TechArena counts 104) |
| Cloud-based systems | 57 (37%, up from 29% in v6.0) |
| Systems with 1,000+ accelerators | 12; largest 10,240 |
| Benchmarks | 7 (2 new) |
The two new workloads are Qwen3.5-397B GRPO — post-training with Group Relative Policy Optimization that teaches the model to work through software-engineering tasks such as tool calling — and DLRMv4-HSTU, a recommender built on an attention-based architecture that replaces the older DLRM-dcnv2 (so recommender results aren't comparable with earlier rounds). MLCommons says the working group spent about two years building the GRPO test.
TechArena notes why it's different: in RL, the model generates its own attempts, they are graded in sandboxed tests on ordinary CPUs, and the weights are updated — so the benchmark measures the whole system, not just the accelerator. It counted 29 GRPO submissions. TechArena also reports that every result landed in the closed division's "available" category, meaning the hardware can be bought or rented now.
Nvidia: Rubin's first numbers
Nvidia says Vera Rubin NVL72 was fastest on six of seven benchmarks and up to 2x faster per GPU than GB300 NVL72 (Blackwell Ultra), and that it was the only vendor to submit on all seven. Selected figures from its post, time to train in minutes:
| Benchmark | GPUs | Vera Rubin | GB300 | Nvidia's per-GPU speedup |
|---|---|---|---|---|
| DeepSeek-V3 671B | 1,024 | 4.14 | 8.03 | 2x |
| Qwen3.5-397B GRPO | 1,024 | 10.44 | 13.78 | 1.4x (claimed) |
| gpt-oss 20B | 64 | 11.213 | 18.24 | 1.6x |
| Llama 3.1 8B | 72 | 7.162 | 11.541 | 1.6x |
| DLRMv4 | 128 | 2.829 | 3.871 | 1.4x |
Two caveats from reading the post: its table lists the GRPO run at 64 GPUs while the text says 1,024, and 13.78 ÷ 10.44 works out to roughly 1.3x rather than the 1.4x shown. Nvidia also published faster "reruns" on newer software that MLCommons has not reviewed — treat those as claims, not results. The fastest absolute DeepSeek-V3 time was still a Blackwell system: 1.38 minutes on 10,240 GB300 GPUs. Nvidia says its platforms made up 80% of closed-division submissions.
AMD: MI355X against Blackwell
AMD submitted MI355X and MI350X on five of the seven benchmarks (not DeepSeek-V3 or GRPO), at 8 to 128 GPUs. In its write-up, AMD claims the MI355X beat Nvidia's B200 on every benchmark with a direct comparison (about 12% on the two Llama tests) while trailing B300 by 4.7–7.4% on Llama. On gpt-oss 20B at eight GPUs it posted 56.18 minutes against 52.90 for a GB300 system. Crusoe, Dell, HPE, MiTAC, Oracle and Supermicro also submitted on AMD hardware, and independents KRAI and tiny corp ran it too.
Google: Ironwood shows up, but read the results yourself
Ironwood (TPU7x) posted its first available-system results, and TechArena says an 8,192-chip Ironwood system was among the three largest DeepSeek-V3 submissions. Neither MLCommons' release nor the coverage we read gives Ironwood timings, so we can't say how it compares; the figures are in MLCommons' results tables.
How to read this
- Vendor blogs pick their comparisons. Nvidia compares Rubin with its own previous generation; AMD compares MI355X with the Nvidia parts it beats or nearly matches. Check the official results tables for like-for-like scale and software.
- Software keeps moving the goalposts. TechArena reports best eight-accelerator times improved 1.75x–2.03x on three benchmarks versus the last round, and AI-assisted kernel tuning appeared for the first time (Firmus cut a GB300 gpt-oss run about 24%).
- Missing entries matter. AMD skipped the two largest LLM workloads, and no Ironwood timings were highlighted. Absence isn't a loss, but it means there's no comparison.
Who should care: anyone renting or buying training capacity for RL fine-tuning now has a public benchmark close to that workload. What to watch: whether AMD and Google submit on the GRPO workload next round, and whether Nvidia's unreviewed reruns hold up once they go through MLCommons review.
Sources
- MLCommons — MLCommons Releases MLPerf Training v6.1 Results (GlobeNewswire) globenewswire.com
- MLCommons — MLPerf Training v6.1 Results: Record Participation mlcommons.org
- NVIDIA Technical Blog — Vera Rubin NVL72 Achieves up to 2x Blackwell Ultra Performance in MLPerf Training v6.1 developer.nvidia.com
- AMD ROCm Blogs — Technical Deep Dive into AMD MLPerf Training v6.1 Results rocm.blogs.amd.com
- TechArena — In Record Round, MLPerf Training 6.1 Reaches Into Post-Training techarena.ai