NVIDIA’s new Vera Rubin NVL72 platform recorded the shortest training time in six of seven tasks in the MLPerf Training v6.1 benchmark suite. In the seventh, DeepSeek-V3 671B, a larger NVIDIA GB300 NVL72 system with 10,240 GPUs produced the fastest result of the round. NVIDIA therefore held the best time across all seven tasks.
At matched GPU counts, Vera Rubin delivered 1.4 to 2 times the performance per GPU of an optimized GB300 NVL72 baseline. NVIDIA said its platforms were the only ones to submit results for all seven tests, and the new system completed the entire suite on its debut. The MLPerf data cited by the company were retrieved on November 16, 2026.
Performance across different models
The closed-division comparison covered language models, image generation, recommendation and reinforcement learning. At matched GPU counts, Vera Rubin’s advantage was:
- DeepSeek-V3 671B: 2 times the performance with 1,024 GPUs.
- Qwen3.5 397B GRPO reinforcement learning: 1.4 times the performance with 64 GPUs.
- gpt-oss 20B: 1.6 times the performance with 64 GPUs.
- Llama 2 70B LoRA: 1.5 times the performance with 72 GPUs.
- Llama 3.1 8B: 1.6 times the performance with 72 GPUs.
- FLUX.1: 1.5 times the performance with 72 GPUs.
- DLRMv4: 1.4 times the performance with 128 GPUs.
The largest difference came on DeepSeek-V3, a 671-billion-parameter mixture-of-experts model. Parallelizing its experts places substantial communication demands on the system.
Vera Rubin NVL72 completed the DeepSeek-V3 test in 4.14 minutes using 1,024 GPUs across 16 racks; the GB300 configuration with the same GPU count took 8.03 minutes. The fastest absolute time of the round, however, came from GB300: it completed the task in 1.38 minutes using 10,240 GPUs across 160 racks.
A new test for agent post-training
MLPerf Training v6.1 includes its first benchmark for reinforcement-learning post-training of large language models for agent tasks. In the Qwen3.5 397B GRPO workload, the system must repeatedly generate and score responses, then update model weights. That requires substantial data movement between inference and training components.
Vera Rubin finished in 10.44 minutes with 1,024 GPUs, compared with 13.78 minutes for GB300 at the same scale. NVIDIA also demonstrated NVIDIA NeMo RL with two inference backends: TensorRT-LLM completed the task in 10.44 minutes and vLLM in 10.94 minutes.
Simulations for agent tasks are also CPU-intensive. According to NVIDIA, 256 processors in a single Vera CPU rack can support more than 22,500 concurrent reinforcement-learning environments while GPUs handle generation and weight updates.
Networking and scaling
Within a rack, Vera Rubin NVL72 connects 72 GPUs and 36 CPUs through a copper NVLink 6 interconnect. NVIDIA says it provides 3.6 TB/s per GPU and 260 TB/s per rack. Between racks, ConnectX-9 SuperNIC network adapters connect systems over Quantum-X800 InfiniBand or Spectrum-X Ethernet; adaptive routing and congestion control manage traffic.
NVIDIA said Vera Rubin maintained more than 97% scaling efficiency on the DeepSeek-V3 test when expanding from 256 to 1,024 GPUs, or from four to 16 racks. GB300 demonstrated scaling on a larger cluster with its 10,240-GPU result.
Software optimization and unverified reruns
For the official runs, NVIDIA NeMo Framework 26.09 supported the pretraining tests, NVIDIA NeMo RL the reinforcement-learning test, and NVIDIA PyTorch releases the FLUX.1 and DLRMv4 workloads. After the v6.1 submissions closed, the company also reran three tests with newer software on the same hardware and configurations.
The reruns were not verified by MLCommons Association or reviewed under MLPerf, and their measurement methods or implementations may differ from the requirements for verified results. The official and rerun times were:
- Llama 3.1 8B, 72 GPUs: GB300 fell from an official 11.541 minutes to 9.211 minutes, and Vera Rubin from 7.162 to 6.140 minutes.
- gpt-oss 20B, 64 GPUs: GB300 changed from 18.24 to 13.151 minutes, and Vera Rubin from 11.213 to 8.080 minutes.
- DLRMv4, 128 GPUs: GB300 fell from 3.871 to 2.940 minutes, and Vera Rubin from 2.829 to 1.628 minutes.
NVIDIA also described an external optimization example. Thinking Machines Lab developed custom kernels and an NVFP4 training method for its Inkling 2 Small model on GB300 NVL72. Compared with full MXFP8 training, pretraining and fine-tuning were 8% faster and reinforcement learning was 20% faster, with minimal impact on accuracy.
Partner results
NVIDIA said 80% of all closed-division submissions ran on its platforms, with participation from 20 partner organizations. According to the company, systems are generally available through several cloud infrastructure partners.
Azure’s 2,560-node, 10,240-GPU GB200 NVL72 submission tied NVIDIA’s GB300 cluster of the same GPU count for the largest cluster in the round. Firmus used AI-assisted kernel generation, automated performance testing and mathematical validation to produce the round’s fastest eight-GPU GB300 results: 52.90 minutes for GPT-OSS 20B and 52.40 minutes for Llama 3.1 8B.



