AivexaNewsSearch
AI news for builders and product teamsChecked every hour

NVIDIA Vera Rubin NVL72 Achieves up to 2x Blackwell Ultra Performance in MLPerf Training v6.1

Collected Nov 16, 2026

NVIDIA's Vera Rubin NVL72 platform made its first appearance in MLPerf Training v6.1 and posted the fastest time to train on six of the seven benchmarks in the suite, according to results NVIDIA published from the round. The company says the new architecture delivers up to 2x more performance per GPU than a highly optimized GB300 NVL72 baseline. NVIDIA platforms were also the only ones to submit across all seven benchmarks, and Vera Rubin completed the full suite in its inaugural run.

The headline number comes from DeepSeek-V3 671B, a 671-billion-parameter, token-dropless Mixture-of-Experts model and the most demanding workload in the set. Vera Rubin NVL72 trained it in 4.14 minutes on 1,024 GPUs across 16 racks; GB300 NVL72 ran the identical 1,024-GPU configuration in 8.03 minutes, close to a 2x per-GPU gain. NVIDIA attributes the size of that gap to the model's communication pattern: expert parallelism turns each iteration into bursts of all-to-all traffic, so fabric efficiency shows up directly in time to train. Across the rest of the suite the platform was consistently 1.4x to 1.6x faster, and the workload mix spans natural language, image generation, recommendation and reinforcement learning.

The absolute fastest result of the round went the other way. On DeepSeek-V3, GB300 NVL72 completed the job in 1.38 minutes spread across 10,240 GPUs, or 160 racks, each its own NVLink domain. That is one training job running past the scale of a single rack, and it demonstrates the scale-out fabric holding up. NVIDIA is deliberately presenting the two platforms as complementary: Vera Rubin for performance per GPU, Blackwell and Blackwell Ultra for the scale record. Azure's 2,560-node, 10,240-GPU GB200 NVL72 submission tied NVIDIA's 10,240-GPU GB300 NVL72 run for the largest clusters in the round.

MLPerf Training v6.1 adds its first agentic reinforcement learning post-training benchmark, Qwen 3.5 397B GRPO, reflecting a shift in how models are finished for agentic use. The workload is punishing for system design because it loops continuously through generating rollouts, scoring them and updating weights, which forces constant large weight migrations between inference and training engines. Vera Rubin NVL72 posted the fastest time at 10.44 minutes on 1,024 GPUs, against 13.78 minutes for GB300 NVL72 at the same scale. NVIDIA also submitted the benchmark on two inference backends, NeMo RL with TensorRT-LLM at 10.44 minutes and with vLLM at 10.94 minutes, a sign that the software stack is already broad rather than tuned around a single path.

The CPU side matters more in this workload than in classic pretraining. RL rollouts run in simulated environments and execute tool calls, code or API validation, which is intensely CPU-bound. NVIDIA says a single Vera CPU rack sustained more than 22,500 concurrent RL environments across 256 CPUs while the GPUs handled token generation and weight updates, keeping the accelerators fed with feedback.

Networking underpins every result. Inside a rack, Vera Rubin NVL72 links 72 GPUs and 36 CPUs over a copper NVLink 6 spine delivering 3.6 TB/s per GPU and 260 TB/s per rack, which keeps MoE expert dispatch off the critical path. Beyond the rack the platform scales out through ConnectX-9 SuperNICs over Quantum-X800 InfiniBand or Spectrum-X Ethernet, with adaptive routing and congestion control aimed at bursty all-to-all traffic. Vera Rubin NVL72 held above 97% scaling efficiency going from 256 to 1,024 GPUs, four to 16 racks, on DeepSeek-V3. In practice, a consistent 72-GPU domain means teams can expect per-generation jumps in compute and bandwidth without changing how they operate their clusters.

NVIDIA says the submitted results used the software customers actually receive: NeMo Framework 26.09 for pretraining benchmarks, NeMo RL for the RL benchmark, and NVIDIA PyTorch releases for FLUX.1 and DLRMv4. Since submissions closed, the company has re-run several benchmarks on identical hardware with a newer build and reports up to 2.4x faster time to train on Vera Rubin NVL72 versus the v6.1 GB300 NVL72 submissions. Those reruns have not been verified by MLCommons and did not go through MLPerf review, so they should be read as a directional signal about ongoing optimization rather than a benchmarked result.

The round also shows what customers can extract themselves. Thinking Machines Lab built an NVFP4 training recipe with custom kernels for its Inkling 2 Small model on GB300 NVL72, using NVFP4 for forward-pass MoE GEMMs and MXFP8 for gradients; NVIDIA reports that this sped up pretraining and fine-tuning by 8% and reinforcement learning by 20% against full MXFP8 training, with minimal accuracy impact. Wider ecosystem participation was strong: 80% of closed-division submissions ran on NVIDIA platforms across 20 partner organizations. On the node level, Firmus used AI-driven kernel generation with automated performance testing and mathematical validation to post the fastest GB300 8-GPU results of the round, training GPT-OSS 20B in 52.90 minutes and Llama 3.1 8B in 52.40 minutes. Vera Rubin NVL72 systems are generally available through cloud infrastructure partners.

Why it matters: Teams choosing training infrastructure now have a verified, first-round data point that NVIDIA's next architecture improves per-GPU throughput by roughly 1.4x to 2x on identical GPU counts, with the biggest gain on the most communication-heavy model. Because the gains appear across every benchmark, including the new GRPO post-training workload that mirrors agent development loops, developers can expect faster experiment cycles without reworking their cluster setup or software path. The caveat worth tracking is that verified numbers and vendor reruns on newer software diverge, so procurement and capacity planning should lean on the submitted results while treating post-round optimization claims as upside.

Read at NVIDIA Developer Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

As AI agents tackle increasingly complex workflows, reducing time-to-train is critical for rapidly iterating on both models and the agents they power. MLPerf...