Nvidia unveils Vera Rubin platform with 10x token efficiency

Nvidia has unveiled new performance benchmarks and architectural details for its upcoming Vera Rubin AI platform, showcasing significant efficiency gains as the company prepares for its release later this year. The announcements, made during a technical workshop at the company's Santa Clara headquarters last week, highlight Nvidia's push to supply complete AI systems, including both graphics processing units and central processing units, as the industry shifts toward complex, agentic AI workloads.

The Vera Rubin platform is Nvidia's successor to its Grace Blackwell system and is designed as a rack-scale solution. A single Vera Rubin NVL72 system combines 36 Vera CPUs with 72 Rubin GPUs. Early production tests conducted by cloud provider CoreWeave demonstrated a tenfold increase in tokens processed per watt when running the DeepSeek R1 model on the new platform compared to the previous generation. Nvidia stated these gains were validated across over 250,000 configurations and 1.4 million GPU hours of testing.

A key focus of the new platform is performance for AI agents, which require CPUs to orchestrate data flows and software tasks. Nvidia's Vera CPU, built on a custom...