NVIDIA's next-generation Vera Rubin AI computing platform has delivered up to twice the performance per GPU of its GB300 NVL72 predecessor in several artificial intelligence training benchmarks, according to results released on November 16 by MLCommons.
The results mark the first appearance of the Vera Rubin architecture in MLPerf Training v6.1, an industry benchmark designed to measure how quickly computing systems can train machine-learning models to a specified level of accuracy.
NVIDIA's Vera Rubin NVL72 platform delivered leading performance across six of the seven training benchmarks in its debut, according to the company's analysis of the results.
The figures are significant because AI development increasingly requires large computing systems capable of training sophisticated models and supporting complex AI agents.
However, the headline performance improvement needs to be understood in context. Benchmark results compare specific hardware configurations and workloads. They do not mean that every AI application will automatically run twice as fast.
What MLPerf measures
MLPerf is an industry benchmark programme operated by MLCommons, a consortium that develops standardised measurements for artificial intelligence hardware and software.
Its training tests evaluate how quickly a system can train different types of machine-learning models while meeting defined accuracy requirements.
The results allow developers and organisations to compare computing systems using common workloads.
The latest Training v6.1 round includes tests covering large language models, recommendation systems, image generation and reinforcement learning post-training.
MLCommons reported 153 submitted systems and 250 test results from 24 organisations.
The round also included 57 cloud-based systems, representing 37 per cent of all submitted systems.
This broad participation is important because AI computing performance depends on more than the processor itself. Hardware configuration, networking, software optimisation and system scale all influence results.
Vera Rubin's performance results
NVIDIA reported that Vera Rubin NVL72 achieved up to twice the performance per GPU of an optimised GB300 NVL72 system on the DeepSeek-V3 671B benchmark.
The comparison used 1,024 GPUs.
The platform also delivered improvements across several other workloads.
In the published results, Vera Rubin achieved approximately 1.4 times the performance of GB300 on the Qwen3.5 397B reinforcement-learning benchmark and DLRMv4 recommendation workload.
It achieved around 1.5 times performance on Llama 2 70B LoRA and FLUX.1, while the Llama 3.1 8B and gpt-oss 20B comparisons reached approximately 1.6 times.
These are per-GPU performance comparisons between specific submitted systems, not universal application speed guarantees.
NVIDIA said Vera Rubin completed the full benchmark suite in its first appearance.
For AI developers, the range of workloads matters because training systems are increasingly expected to support different model architectures and development stages.
Why reinforcement learning matters
One of the important additions to MLPerf Training v6.1 is a large-scale post-training benchmark using Qwen3.5 397B and Group Relative Policy Optimization.
Post-training refers to additional model development after initial pretraining.
It can be used to improve a model's ability to follow instructions, use tools and solve particular tasks.
Reinforcement learning is one approach in which models improve their behaviour using feedback associated with task performance.
The benchmark focuses on the kind of training that supports more advanced agent-based capabilities, including tool use and software engineering tasks.
Its inclusion reflects a shift in AI development. Training a large language model is no longer the only major computing challenge. Companies also need infrastructure to refine models and improve their ability to perform multi-step tasks.
More organisations are testing AI hardware
The latest MLPerf round demonstrates that the AI infrastructure market is becoming more diverse.
MLCommons reported submissions from 24 organisations, including NVIDIA, AMD, Google, Microsoft Azure, Dell, Oracle and several cloud computing providers.
Twelve submitted systems used at least 1,000 accelerators.
The largest system in the round used 10,240 accelerators.
The benchmark therefore reflects a market in which computing performance increasingly depends on the ability to connect large numbers of processors into coordinated systems.
For cloud providers, this can influence infrastructure investment decisions.
For AI developers, faster training can reduce the time needed to experiment with models and potentially improve the economics of developing new systems.
However, purchasing decisions also depend on power consumption, hardware availability, software compatibility, system reliability and total operating costs.
A faster benchmark result alone does not establish which platform offers the best overall value.
The importance of independent benchmarking
NVIDIA's published results are useful for understanding the company's platform, but it is important to distinguish the manufacturer's interpretation from the independent benchmark programme.
MLCommons provides the wider results and explains the methodology behind the testing.
NVIDIA also published additional benchmark reruns using a newer software stack after the official submissions had been finalised.
The company explicitly identified these additional measurements as unverified results that had not gone through MLCommons review.
Those figures should not be presented as official MLPerf results.
The distinction is important because benchmark comparisons are only useful when the measurement method, hardware configuration and workload are clearly identified.
What the results mean for AI development
The Vera Rubin debut provides an indication of how quickly computing infrastructure is evolving to support more demanding AI workloads.
The addition of large-scale post-training benchmarks also shows how the industry is adapting its performance measurements to newer development methods.
The next challenge is turning benchmark gains into practical improvements for users.
That will depend on how efficiently the hardware performs in real deployments, how much electricity the systems consume and how effectively software developers can use the available computing capacity.
Organisations adopting AI infrastructure will need to assess these factors alongside raw performance.
For NVIDIA, the new benchmark results provide a strong early demonstration of the Vera Rubin platform.
For the wider industry, the more important development may be the increasing availability of standardised comparisons across different AI workloads and hardware providers.





Comments
0