Skip to content
0MESubscribe
AI16 Nov 20262 min read

Nvidia Vera Rubin delivers 10x throughput per watt in debut MLPerf AI tests

Nvidia's Vera Rubin hardware swept the latest MLPerf Training tests and showed a 10x efficiency leap on agentic AI workloads compared to the previous generation.

Listen to this storyAbout 3 min, read by your device
Picture made with AI

Nvidia revealed the first MLPerf benchmarkBenchmarkA standard test used to compare AI models or chips. Open in glossary results for its next-generation Vera Rubin computing platform. The hardware took top spots across all MLPerf TrainingTrainingTeaching an AI model by showing it very large amounts of data. Open in glossary v6.1 tests and demonstrated large efficiency gains in AI inferenceInferenceRunning a trained AI model to get an answer. Open in glossary workloads.

Training Gains and Near-Linear Scaling

The Vera Rubin NVL72 system delivered up to twice the performance per graphics processing unitGraphics Processing UnitA specialized electronic circuit designed to rapidly manipulate and alter memory to accelerate the creation of images. Open in glossary compared to Nvidia’s earlier GB300 NVL72 setup. Nvidia stood out as the only hardware maker submitting results for every test in the training round. These workloads covered image generation, recommendation systems, large language models and post-training tasks using reinforcement learning.

Nvidia also tested the hardware at scale using the DeepSeek-V3 benchmark. Vera Rubin NVL72 systems scaled from 256 to 1,024 GPUs with roughly 97 percent efficiency. Quadrupling the cluster size reduced training time from 16 minutes to about four minutes. For larger setups, Nvidia and cloud provider Microsoft Azure both ran configurations using 10,240 GPUs based on the GB300 and GB200 architectures, respectively.

A 10x Boost in Inference Power Efficiency

Beyond model training, Nvidia submitted initial figures to MLPerf Endpoints, a suite measuring real-world agentic workflows such as code generation and enterprise automation. These tasks often require an agent to reason, access external tools and verify answers, consuming significant processing power and generating millions of data tokens.

In preliminary testing using the open-source Kimi K3 model, Vera Rubin NVL72 posted more than 10 times the throughput per wattThroughput per wattA ratio showing how much digital work a computing system finishes for each unit of electrical energy it uses. Open in glossary compared to the GB300 NVL72 at matching response speeds. The Rubin cluster achieved 41,414 tokens per second across 1,024 concurrent queries while consuming 100.89 kilowatts. This power metric is critical for modern data centers facing severe electrical capacity limits.

Twenty partner companies submitted results using Nvidia hardware for the benchmark cycle, including Dell Technologies, Google Cloud, Oracle and Supermicro. Nvidia noted that Rubin’s performance should improve further over time as it updates its underlying CUDACUDAA computing system that lets software use graphics chips for general calculation tasks. Open in glossary software stack.

How does this story land?

One tap per reader.

Quick poll

What is the biggest bottleneck for future data centers?

One tap. No sign-up.

Ask this story

Still have a question?

Answers use only what is in this story. They are written by AI and can be wrong.

Sources

  1. - Archives Page 1 | NVIDIA Blog, NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance
Up next in AIApple begins 200,000-person study on medical AI answers2 min read

Keep reading

All AI stories