Nvidia shares new benchmarks for its Vera Rubin CPU and GPU system

In short: Nvidia is promoting new performance claims for its upcoming Vera Rubin system, which bundles its own CPUs and GPUs for AI data centers.

What happened

Nvidia held a technical workshop in Santa Clara, California, where executives shared new benchmark results for Vera Rubin, its next chip platform after Grace Blackwell. The company said Vera Rubin will ship in the second half of this year, with early customers including Microsoft, OpenAI, and Oracle.

Vera Rubin is designed as a combined setup of GPUs and CPUs. GPUs are the chips that do a lot of the heavy lifting for AI, and CPUs are the chips that coordinate tasks and move data around (like a manager keeping work flowing). Nvidia says one Vera Rubin NVL72 system includes 72 Rubin GPUs paired with 36 Vera CPUs.

Nvidia also said the new racks are easier to install and maintain. It described them as more “plug-and-play,” with fewer cables, “hot-swappable” parts (meaning you can replace some parts without turning everything off), and full liquid cooling. Nvidia executives said OpenAI already has one Vera Rubin rack in use.

On performance, Nvidia claimed the system can process 10 times as many “tokens per watt” as Grace Blackwell. Tokens are small chunks of text that an AI system reads and writes, so this is a way of measuring how much text work the system can do for the same amount of electricity. Nvidia also said its Vera CPU is faster for “agentic” AI tasks, meaning AI tools that can plan steps and carry out actions, though Wired noted Nvidia compared against slightly older rival chips from AMD and Intel.

Why it matters

Big AI systems run in data centers, which are basically warehouses full of computers. If Nvidia can sell not just the GPU “engine” but also the CPU “traffic controller,” it can become a one-stop supplier for more of what powers AI services people use every day.

Source: Wired