Vera Rubin Sweeps All Categories in MLPerf Debut, Doubling Per-GPU Compute to Define the 2027 Compute Baseline

In its first MLPerf Training v6.1 submission, NVIDIA's Vera Rubin NVL72 was the only platform to submit results across all benchmark categories, delivering

On November 16, 2026, MLCommons released the MLPerf Training v6.1 results, with 24 organizations submitting 250 test results across 153 systems. This benchmark included reinforcement learning post-training in official testing for the first time. NVIDIA's Vera Rubin NVL72 was the only platform to submit results in every benchmark category—covering large language model training, image generation, recommendation systems, and reinforcement learning post-training—delivering up to 2x higher per-GPU training performance than GB300 NVL72.

Two Numbers: 2x and 97%

The core figures disclosed in NVIDIA's official blog are: at the same GPU count, Vera Rubin NVL72 delivers up to twice the per-GPU training performance of GB300 NVL72; on the DeepSeek-V3 benchmark, when the system scaled from 256 GPUs to 1,024 GPUs, overall efficiency remained at about 97%, compressing training time from 16 minutes to about 4 minutes.

97% scaling efficiency means that when the GPU count quadruples, training time is reduced nearly linearly. NVIDIA's explanation is that NVLink provides intra-rack GPU interconnect, while Quantum-X800 InfiniBand and Spectrum-X Ethernet handle inter-rack communication, with coordinated software and hardware optimization achieving this result.

NVIDIA also demonstrated extreme scale in this round: a GB300 system ran the largest-scale test of the round with 10,240 GPUs, and Microsoft Azure followed with the same scale of 10,240 GB200 GPUs.

First Cloud Example: CoreWeave's Multi-Rack Deployment Data

CoreWeave submitted production results for Vera Rubin across multiple racks. With 256 Vera Rubin GPUs deployed across multiple racks, full training of the 671B-parameter DeepSeek-V3 took 17.65 minutes, 2.7x faster than GB200 NVL72 at the same GPU count. The single-rack Llama 3.1 8B test (72 GPUs) took 7.16 minutes, about 1.9x better than the GB200 baseline (13.70 minutes). CoreWeave used a Spectrum-X Ethernet AI interconnect fabric with ConnectX-9 SuperNICs.

Behind the "10x" Agentic Inference Figure Lies a Footnote

In its debut in MLPerf Endpoints, Vera Rubin delivered more than 10x higher throughput per watt than GB300 NVL72 on the Kimi K3 agentic workload with 32 Vera Rubin accelerators, consuming only 100.89 kW at equivalent response speed and achieving 41,414.2 tokens/second throughput.

NVIDIA's official blog noted that the Kimi K3 result is provisional and pending peer review validation by the MLCommons association, and the result may change upon review.

Google Ironwood and AMD: Two Other Teams on the Same Battlefield

Google's TPUv7 (Ironwood) also made its debut in the same round, with 8,192 Ironwood TPUs completing the DeepSeek-V3 benchmark in 1.70 minutes while maintaining 96% scaling efficiency. Google included agentic reinforcement learning coding tasks in testing for the first time; 640 Ironwood TPUs completed the task in 18.02 minutes, with chip idle time below 3% during parallel code generation and learning.

AMD's MI355X and MI350X (CDNA 4 architecture) delivered a 3.1x generational improvement on the FLUX.1 image generation benchmark—training time dropped from 92.36 minutes to 29.58 minutes in a 64-GPU configuration.

NVIDIA is the only participant with comparable results in every benchmark category, meaning its "clean sweep" partly stems from competitors' selective participation. In the compute market, full coverage indicates that the platform's generality has reached the point where it can deliver results on all mainstream training tasks.

What the 2027 Compute Baseline Means

MLPerf v6.1 is the first version to include reinforcement learning post-training (represented by the Qwen3.5-397B GRPO task) in the official benchmark. As large-model training evolves from "single pretraining" to "continuous post-training + reinforcement learning alignment," compute demand is shifting from single large-batch training toward more frequent, smaller-scale reinforcement learning iterations.

Vera Rubin's 97% scaling efficiency and CoreWeave's 2.7x figure form a quantitative anchor for 2027 data center procurement. At the SC26 supercomputing conference, NVIDIA announced Rubin HPC for scientific computing, featuring more than 50,000 CUDA cores and expected to launch in 2028.

When Benchmarks Become Market Power

MLPerf is one of the most frequently cited third-party sources of numbers in data center procurement negotiations. Vera Rubin's 2x per-GPU improvement and 97% scaling efficiency under full coverage will directly appear in the technical specifications for 2027 cloud provider and enterprise compute procurement.

Google Ironwood's scaled performance is also not to be ignored, and AMD's generational gains are accelerating. But based on the publicly available benchmark data from November 16, NVIDIA has used Vera Rubin to give a numeric answer to "what the next-generation GPU should be able to do," and it has put that answer through an independent organization's evaluation process. In the 2027 compute competition, this number will be cited repeatedly.