Model Comparisons

NVIDIA Vera Rubin NVL72 Achieves Up to Twice the Performance of Blackwell Ultra in MLPerf Training v6.1

Vera Rubin NVL72 leads six MLPerf v6.1 benchmarks, up to 2x Blackwell; 10,240 GPUs set the fastest DeepSeek-V3 time, covering all seven workloads.

NVIDIA Vera Rubin NVL72 Achieves Up to Twice the Performance of Blackwell Ultra in MLPerf Training v6.1

Key Takeaways

  • NVIDIA Vera Rubin NVL72 led six of seven MLPerf Training v6.1 closed-division benchmarks, delivering up to 2x faster time-to-train per GPU versus optimized GB300 NVL72.
  • GB300 NVL72 set the fastest absolute DeepSeek-V3 671B result at 1.38 minutes on 10,240 GPUs, while Vera Rubin maintained above 97% scaling efficiency from 256 to 1,024 GPUs.
  • Vera Rubin NVL72 also won the new Qwen3.5 397B GRPO reinforcement-learning benchmark at 10.44 minutes, showing software maturity across NeMo RL, TensorRT-LLM, and vLLM backends.
  • The results reflect co-designed NVLink, CPU simulation capacity, ConnectX-9 networking, and mature CUDA software supporting broad AI workloads from pretraining to agentic post-training.

Meta description: NVIDIA Vera Rubin NVL72 leads six MLPerf Training v6.1 benchmarks, while GB300 NVL72 records the fastest absolute DeepSeek-V3 671B time on 10,240 GPUs.

NVIDIA Vera Rubin NVL72 Achieves Up to Twice the Performance of Blackwell Ultra in MLPerf Training v6.1

Tags: NVIDIA, MLPerf Training, Vera Rubin, GB300 NVL72, AI agents

Featured image: abstract, friendly illustration of linked server racks and flowing data paths, no real people, no logos, no text.

AI infrastructure teams and developers evaluating large model training platforms can use MLPerf Training v6.1 results to see where NVIDIA's new Vera Rubin architecture improves training speed, scaling, and reinforcement-learning workflows. In the round, the Vera Rubin NVL72 platform recorded the fastest time-to-train on six of seven closed-division benchmarks, delivering up to 2x performance versus a highly optimized NVIDIA GB300 NVL72 baseline. For the remaining DeepSeek-V3 671B workload, NVIDIA took the round's fastest absolute time by running GB300 NVL72 on a 10,240-GPU cluster. Together, the two platforms cover every benchmark in v6.1.

A debut with broad coverage

NVIDIA platforms were the only ones to submit results across all seven benchmarks in MLPerf Training v6.1. The Vera Rubin NVL72 platform completed the entire suite in its first run, which matters for teams that need one system to support pretraining, fine-tuning, image generation, recommendation, and reinforcement-learning post-training. The company said the CUDA software stack was already highly mature, so the early results reflect day-one performance across the model pipeline rather than a single optimized laboratory peak.

The benchmark table below shows Vera Rubin NVL72 performance per GPU relative to optimized GB300 NVL72 at the same GPU count in the closed division.

Model typeBenchmarkGPU countVera Rubin vs GB300Submission ID Vera Rubin / GB300
Mixture-of-ExpertsDeepSeek-V3 671B1,0242x6.1-0116 / 6.1-0030
Mixture-of-ExpertsQwen3.5 397B GRPO reinforcement learning641.4x6.1-0135 / 6.1-0020
Mixture-of-Expertsgpt-oss 20B641.6x6.1-0135 / 6.1-0049
DenseLlama 2 70B LoRA721.5x6.1-0123 / 6.1-0138
DenseLlama 3.1 8B721.6x6.1-0073 / 6.1-0138
DiffusionFLUX.1721.5x6.1-0124 / 6.1-0138
RecommenderDLRMv41281.4x6.1-0129 / 6.1-0038

The pattern is broad: gains range from 1.4x to 1.6x per GPU across most workloads, while the largest gain appears on the largest, most communication-heavy workload. The suite covers natural language, image generation, recommendation, and reinforcement learning, so the improvements are not limited to one model shape or modality.

NVIDIA Vera Rubin NVL72 Achieves Up to Twice the Performance of Blackwell Ultra in MLPerf Training v6.1

DeepSeek-V3 and 10,240-GPU scale

DeepSeek-V3 671B is a 671-billion-parameter token-dropless Mixture-of-Experts model and the suite's most demanding workload. In this benchmark, expert parallelism turns each iteration into a burst of all-to-all communication, so network fabric efficiency shows up directly in training time.

Vera Rubin NVL72 trained DeepSeek-V3 671B in 4.14 minutes using 1,024 GPUs across 16 racks. GB300 NVL72 ran the same 1,024-GPU configuration in 8.03 minutes, a result close to 2x faster per GPU and the largest generational improvement in the round. If time-to-train is halved, a team can run roughly twice as many experiments in a quarter, which is how many model-quality decisions get made.

At even larger scale, GB300 NVL72 produced the fastest absolute result of the round: 1.38 minutes across 10,240 GPUs. That job spanned 160 racks, each its own NVLink domain. The result highlights reliable cluster scale, while Vera Rubin showed its own scaling behavior by maintaining above 97% overall efficiency when moving from 256 to 1,024 GPUs, or from 4 to 16 racks.

Post-training moves into the benchmark suite

MLPerf Training v6.1 added its first agentic large language model reinforcement-learning post-training benchmark, Qwen3.5 397B GRPO. The addition reflects a shift in AI development: post-training is becoming a key differentiator for agentic accuracy. The workload is demanding because it repeatedly generates rollouts, scores them, and updates weights, forcing large weight migrations between inference and training engines.

Vera Rubin NVL72 achieved the fastest time for the new benchmark at 10.44 minutes on 1,024 GPUs, submission 6.1-0117. GB300 NVL72 recorded 13.78 minutes at the same scale, submission 6.1-0031. NVIDIA also submitted the benchmark on two inference backends: NVIDIA NeMo RL with TensorRT-LLM at 10.44 minutes and with vLLM at 10.94 minutes, submission 6.1-0118. Having two viable backends on a brand-new benchmark says as much about software maturity as about hardware.

NVIDIA Vera Rubin NVL72 Achieves Up to Twice the Performance of Blackwell Ultra in MLPerf Training v6.1

Training agents is not only a GPU task. Each rollout can require a simulated environment plus tool calls, code, or API validations, which is intensely CPU-bound. The NVIDIA Vera CPU is part of the architecture, sustaining more than 22,500 concurrent reinforcement-learning environments across 256 CPUs in a single Vera CPU rack. GPUs handle large token generation and rapid weight updates, while Vera CPUs simulate environments so the accelerators receive continuous feedback.

Co-designed networking and software

Every result in the round is also a networking result. Modern AI training asks thousands of processors to behave as one system. Inside the rack, Vera Rubin NVL72 connects 72 GPUs and 36 CPUs through a copper NVLink 6 spine. The company said the domain delivers 3.6 TB/s per GPU and 260 TB/s per rack, keeping Mixture-of-Experts expert dispatch off the critical path. Outside the rack, the platform scales with ConnectX-9 SuperNICs over Quantum-X800 InfiniBand or Spectrum-X Ethernet. Both fabrics use adaptive routing and congestion control to keep bursty all-to-all traffic from blocking compute.

The results came from software customers receive: NVIDIA NeMo Framework 26.09 for pretraining benchmarks, NVIDIA NeMo RL for the reinforcement-learning benchmark, and NVIDIA PyTorch releases for FLUX.1 and DLRMv4. The optimizations were described in the earlier MLPerf Training v6.0 round. NVIDIA also said it has re-run several benchmarks since v6.1 submissions were finalized on the same hardware and configuration with a newer software build. Those reruns are unverified results, not MLCommons submissions, but they show potential for continued performance gains after deployment. The company reported up to 2.4x faster time-to-train for Vera Rubin NVL72 versus the MLPerf 6.1 GB300 NVL72 submission, and up to 1.4x faster versus the MLPerf 6.1 Vera Rubin submission.

The CUDA platform also gives customers room to optimize their own kernels. Thinking Machines Lab developed an NVFP4 training recipe and custom kernels for its Inkling 2 Small model on GB300 NVL72, using NVFP4 for forward-pass Mixture-of-Experts matrix multiplication operations and MXFP8 for gradients. Compared with full MXFP8 training, it sped up pretraining and fine-tuning by 8% and reinforcement learning by 20%, with minimal accuracy impact.

Availability and what follows

NVIDIA said 80 percent of all closed-division submissions ran on NVIDIA platforms across 20 partner organizations, suggesting the results are reproducible across clouds and original equipment manufacturers, not isolated laboratory peaks. Vera Rubin systems are already generally available across various cloud infrastructure partners, complementing AI Factory deployments around the world. In the round, Vera Rubin led performance per GPU across all seven workloads, while Blackwell and Blackwell Ultra maintained the scale record. Azure's 2,560-node, 10,240-GPU GB200 NVL72 submission tied the NVIDIA 10,240-GPU GB300 NVL72 run for the largest clusters in this round.

Partners are also optimizing smaller systems. Firmus used AI-driven kernel generation, automated performance testing, and mathematical validation to post the fastest GB300 8-GPU result of the round, training GPT-OSS 20B in 52.90 minutes, submission 6.1-0091, and Llama 3.1 8B in 52.40 minutes, submission 6.1-0094. Because these platforms share a continuously optimizing CUDA software stack, initial MLPerf numbers may keep improving before the next round. NVIDIA points readers to the full MLPerf Training v6.1 results and its MLPerf Endpoints v1.0 blog for inference details.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.