NVIDIA’s Vera Rubin NVL72 made its MLPerf Training debut on November 16 and swept all seven benchmarks in the v6.1 round, posting up to 2x per-GPU performance over the GB300 NVL72. The same round quietly introduced something more significant: the first post-training benchmark in MLPerf Training history.
The Numbers
The headline claim holds up on the right workloads. On DeepSeek-V3 671B — the most demanding benchmark — a 1,024-GPU Vera Rubin cluster finished in 4.14 minutes against 8.03 minutes for an equivalently sized GB300 system. That is a clean 2x. Across the other six benchmarks, gains run 1.4x to 1.6x.
| Benchmark | Vera Rubin vs GB300 |
|---|---|
| DeepSeek-V3 671B (MoE) | ~2x |
| GPT-OSS 20B | 1.6x |
| Llama 3.1 8B | 1.6x |
| FLUX.1 (Diffusion) | 1.5x |
| Llama 2 70B LoRA | 1.5x |
| DLRMv4 (Recommender) | 1.4x |
The architecture behind this is a 72-GPU, 36-CPU rack connected by NVLink 6 at 3.6 TB/s bidirectional per GPU and 260 TB/s across the entire rack backplane. That interconnect density is why Vera Rubin hit 97% scaling efficiency when expanding the DeepSeek-V3 workload from 256 to 1,024 GPUs — a figure that matters far more to actual cluster operators than peak single-node throughput.
The More Important Story: MLPerf’s First Post-Training Benchmark
Generational GPU gains are expected. What is new is the Qwen3.5-397B GRPO benchmark — Group Relative Policy Optimization — which measures how fast a system can apply reinforcement learning to teach a large model software-engineering tasks including tool calling.
This is the first time MLPerf Training has tested post-training workloads. That is a significant shift. Pretraining throughput is what made last decade’s GPU benchmarks relevant; post-training RL is where the real product work happens now. Getting a base model to follow instructions, use tools, and reason step by step requires GRPO-style fine-tuning, not more pretraining tokens.
Vera Rubin completed the GRPO benchmark in 10.44 minutes on 1,024 GPUs. The Vera CPU in the same rack simultaneously sustained over 22,500 concurrent RL environments — the CPU-side workload that GRPO demands at scale is as important as GPU throughput here.
Google’s Ironwood TPU7x completed the same benchmark in 18.02 minutes (640 TPUs), and CoreWeave’s GB200 cluster took 22.78 minutes (256 GPUs). Those numbers are not directly comparable — different cluster sizes make chip-to-chip ranking impossible from these figures — but they confirm that every major vendor is now being tested on post-training, not just pretraining.
Where AMD Stands
AMD submitted five of seven benchmarks using Instinct MI355X hardware, its most comprehensive MLPerf Training submission yet. The results are genuinely competitive with Blackwell: MI355X came within 4-7% of B300 on Llama tasks and was 18.4% faster than B200 on FLUX.1. AMD also logged a 1.19x round-over-round improvement on the same hardware — a pure software gain from ROCm and Primus framework updates, with the first MXFP6 submission in MLPerf history.
What AMD did not submit: the GRPO post-training benchmark. That is the gap worth watching. AMD is now a credible Blackwell-tier competitor. Whether it can close the gap on RL-heavy agentic workloads in future rounds is a different question.
What This Means for Infrastructure Planning
A few practical takeaways for teams planning training infrastructure:
- Vera Rubin is a step-change generation, not an incremental one. The 2x on MoE models and the GRPO performance both point to architectural decisions — HBM4, NVLink 6, Vera CPU — that compound rather than add.
- The CUDA software stack arrived mature. NVIDIA submitted with dual inference backends (TensorRT-LLM and vLLM) for GRPO on day one. That is unusual for a new architecture launch and de-risks adoption.
- Cloud deployment is already here. Azure is running matching 10,240-GPU Vera Rubin clusters. If you need Rubin-class compute today, cloud access is the realistic path.
- The new MLPerf Endpoints v1.0 inference benchmark showed Vera Rubin delivering over 10x throughput per watt versus GB300 on Kimi K3 at a shared 50-token-per-second minimum interactivity threshold.
The Takeaway
The 2x headline is real, but it is the expected result of a full generational jump. The more consequential development is that MLPerf finally added a post-training benchmark. Twenty-four organizations submitted results to a suite that now tests the full training pipeline — pretraining, fine-tuning, and RL-based post-training. That is the benchmark evolution the industry needed. Hardware that cannot handle GRPO at scale will not be able to train the next generation of agentic models efficiently, and now there is a standardized number to prove it.

