Tuesday, November 17, 2026
AI 인프라 · 뉴스 & 분석
홈 › 반도체·하드웨어 › 리포트
반도체·하드웨어 · 리포트

NVIDIA는 Vera Rubin NVL72가 MLPerf Training v6.1에서 Blackwell Ultra 대비 최대 2배의 성능을 달성했다고 밝혔다.

Blackwell Ultra 대비 최대 2배의 학습 성능 향상이라는 주장은 구매자들이 Rubin 기반 설비 투자를 정당화할 수 있는 세대별 비교 기준을 제공한다.
업계 전문지Slicast · 2026년 11월 16일 16:07 UTC · 미국 · 출처: NVIDIA Developer
중요도 65

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 Training v6.1 marks the debut of the NVIDIA Vera Rubin architecture, and these early results demonstrate that its CUDA software stack is already highly mature. In its first appearance, the Vera Rubin NVL72 platform delivered the fastest time to train on six of the seven benchmarks, achieving up to 2x more performance than a highly optimized NVIDIA GB300 NVL72 baseline.

For the seventh benchmark, DeepSeek-V3 671B, NVIDIA captured the absolute fastest time of the round by demonstrating massive scale on a 10,240-GPU GB300 NVL72 cluster. This dual-platform approach deliberately showcases the strengths of both systems. While the GB300 results highlight the massive, reliable cluster scale NVIDIA is deploying today, Vera Rubin demonstrated its own exceptional scaling architecture by maintaining above 97% overall efficiency when scaling from 256 to 1,024 GPUs. Together, NVIDIA platforms hold the fastest time to train on every benchmark in v6.1.

NVIDIA platforms were also the only ones to submit results across all seven benchmarks, with Vera Rubin successfully completing the entire suite in its inaugural run. For developers, this breadth matters as much as any single top-line number. It proves that the performance gains inherited on day one span the entire modeling pipeline, comprehensively supporting everything from large-scale pretraining to the reinforcement learning post-training required for modern AI agents.

## Comprehensive AI performance across all MLPerf training benchmarks

The simplest way to read this round is per benchmark: How much more performance does the new platform deliver per GPU on the same model?

**Table 1. Performance per GPU of Vera Rubin NVL72 relative to GB300 NVL72, MLPerf Training v6.1, closed division**

| Model type | Benchmark | GPU Count | Vera Rubin NVL72 vs. optimized GB300 NVL72 | Submission ID (Vera Rubin / GB300) |

|---|---|---|---|---|

| Mixture-of-Experts | DeepSeek-V3 671B | 1,024 | 2x | 6.1-0116 / 6.1-0030 |

| Mixture-of-Experts | Qwen3.5 397B GRPO (RL) | 64 | 1.4x | 6.1-0135 / 6.1-0020 |

| Mixture-of-Experts | gpt-oss 20B | 64 | 1.6x | 6.1-0135 / 6.1-0049 |

| Dense | Llama 2 70B LoRA | 72 | 1.5x | 6.1-0123 / 6.1-0138 |

| Dense | Llama 3.1 8B | 72 | 1.6x | 6.1-0073 / 6.1-0138 |

| Diffusion | FLUX.1 | 72 | 1.5x | 6.1-0124 / 6.1-0138 |

| Recommender | DLRMv4 | 128 | 1.4x | 6.1-0129 / 6.1-0038 |

*MLPerf Training v6.1 results retrieved from www.mlcommons.org on Nov 16, 2026, from the following entries: 6.1-0116, 6.1-0030, 6.1-0135, 6.1-0020, 6.1-0049, 6.1-0123, 6.1-0138, 6.1-0073, 6.1-0124, 6.1-0129 and 6.1-0038. The MLPerf name and logo are trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use is strictly prohibited. See www.mlcommons.org for more information.*

**Figure 1.** Vera Rubin NVL72 performance relative to GB300 NVL72 at the same GPU count on all seven MLPerf Training v6.1 benchmarks

*MLPerf Training v6.1 results retrieved from www.mlcommons.org on Nov 16, 2026, from the following entries: 6.1-0116, 6.1-0030, 6.1-0135, 6.1-0020, 6.1-0049, 6.1-0123, 6.1-0138, 6.1-0073, 6.1-0124, 6.1-0129 and 6.1-0038. The MLPerf name and logo are trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use is strictly prohibited. See www.mlcommons.org for more information.*

The pattern is clear. The largest gain lands on the largest, most communication-heavy model. While the platform is consistently 1.4x to 1.6x faster across the rest of the suite, the generational advantage widens even further for the massive DeepSeek-V3 benchmark.

The suite covers natural language, image generation, recommendation, and reinforcement learning, so the gains hold across multiple modalities as well as across different model shapes and architectures.

## DeepSeek-V3 671B: 2x per GPU on the hardest workload

DeepSeek-V3 671B is a 671-billion-parameter token-dropless Mixture-of-Experts (MoE) model and the most demanding workload in the suite. Expert parallelism turns every iteration into a burst of all-to-all communication, so fabric efficiency shows up directly in time to train.

Vera Rubin NVL72 trained the benchmark in 4.14 minutes using 1,024 GPUs across 16 racks. GB300 NVL72 ran the identical 1,024-GPU configuration in 8.03 minutes, close to 2x per GPU, and the largest generational improvement in this round. Halving time to train roughly doubles how many experiments a team can run in a quarter, and experiment throughput is what moves model quality.

At a larger scale, GB300 NVL72 produced the fastest absolute result of the round: 1.38 minutes across 10,240 GPUs. That is one training job spread across 160 racks, each its own NVLink domain, and it shows the platform's scale-out fabric holding up long after the job outgrows a single rack.

## Post-training arrives, and Vera Rubin leads it

MLPerf Training v6.1 adds its first agentic LLM reinforcement learning post-training benchmark, Qwen 3.5 397B GRPO, reflecting a fundamental shift in AI development. Post-training has emerged as the critical differentiator for agentic accuracy, making the GRPO workload uniquely demanding. The continuous cycle of generating rollouts, scoring them, and updating weights pushes system architectures to their limits by forcing constant, massive weight migrations between inference and training engines.

Vera Rubin NVL72 achieved the fastest time for this new benchmark with a time of 10.44 minutes at 1,024 GPUs (6.1-0117), compared with 13.78 minutes for GB300 NVL72 (6.1-0031) at the same scale. The Vera Rubin platform also proved its flexibility by submitting on two different inference backends: NVIDIA NeMo RL with TensorRT-LLM at 10.44 minutes and with vLLM at 10.94 minutes (6.1-0118). Shipping two viable inference backends on a benchmark this new says as much about software maturity as it does about silicon.

Teaching these models to observe, reason, and act requires running each rollout in a simulated environment and executing tool calls, code, or API validations. Because driving these simulations is intensely CPU-bound, the NVIDIA Vera CPU is integral to the architecture, sustaining more than 22,500 concurrent RL environments across 256 CPUs in a single Vera CPU rack.

While the GPUs handle massive token generation and rapid weight updates, the Vera CPUs work in lockstep to simulate the environments, ensuring the accelerators are continuously fed with the rich feedback required to perfect agentic workflows.

## NVLink, Quantum-X and Spectrum-X: Co-designed networking for high-efficiency AI scaling

Every result above is also a networking result. Modern AI training requires thousands of processors to act as one, a challenge NVIDIA solves by co-designing the internal rack domain and the external cluster fabric.

Inside the rack, Vera Rubin NVL72 connects 72 GPUs and 36 CPUs through a copper NVLink 6 spine. Delivering 3.6 TB/s per GPU and 260 TB/s per rack, this all-to-all domain keeps MoE expert dispatch off the critical path. Beyond the rack, the platform scales out via ConnectX-9 SuperNICs over Quantum-X800 InfiniBand or Spectrum-X Ethernet. To manage bursty MoE traffic, both fabrics use adaptive routing and congestion control to keep all-to-all communications hidden behind the compute layer.

This cohesive design translates directly to efficiency. On the DeepSeek-V3 benchmark, Vera Rubin NVL72 achieved above 97% scaling efficiency as it scaled from 256 to 1,024 GPUs (4 to 16 racks). GB300 NVL72 showcased the platform's large-scale capabilities, extending to 10,240 GPUs to capture the fastest absolute result of the round.

Because the NVL72 architecture provides a consistent 72-GPU domain, developers see predictable performance jumps every generation without altering their operating model. Each cycle simply delivers more compute and bandwidth across the same consistent fabric.

**Figure 2.** Three generations of NVL72 systems at an identical 1,024-GPU scale

*MLPerf Training v6.1 results retrieved from www.mlcommons.org on Nov 16, 2026, from the following entries: 6.1-0045, 6.1-0030 and 6.1-0116. Ratios are rounded. The MLPerf name and logo are trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use is strictly prohibited. See www.mlcommons.org for more information.*

## Continuous software optimization: Unlocking faster pretraining without hardware upgrades

Vera Rubin's results came from the NVIDIA AI software stack that customers receive: NVIDIA NeMo Framework 26.09 for the pretraining benchmarks, NVIDIA NeMo RL for the reinforcement learning benchmark, and NVIDIA PyTorch releases for FLUX.1 and DLRMv4. The optimizations are the ones described in our MLPerf Training v6.0 round.

The stack also keeps moving after a round closes. Since v6.1 submissions were finalized, we have re-run several of the benchmarks on the

원문 보기
NVIDIA는 Vera Rubin NVL72가 MLPerf Training… · Slicast