Nvidia의 Vera CPU가 출하를 시작했으며, 에이전트형 워크로드를 NVL72 랙에서 분리해 성능을 높이는 전용 '디스크리트' 프로세서 플랫폼도 함께 선보였다.

Nvidia's Vera CPU offerings have begun shipping in "multiple form factors," the company confirmed.

Speaking ahead of the OCP Global Summit in San Jose, Joe DeLaere, Nvidia's product lead for AI infrastructure solutions, said in a press briefing that dedicated Vera racks were "shipping now."

The Vera CPU rack platform leans on Nvidia's MGX modular reference architecture and acts as a dedicated side cabinet for handling agentic workloads. DeLaere described it as a "discrete system" placed next to an NVL72 and separated from the core Vera Rubin platform. The CPUs in an NVL72 are "dealing with other types of orchestration management," leaving not enough cores to perform the tasks needed in agentic workflows.

DeLaere told the press that Vera delivers more than twice the performance on agentic workloads compared with the previous Grace generation, as measured using the SPEC CPU 2026 benchmark. He also cited 1.8-fold performance on Terminal Bench, which he described as measuring agent tasks involving real code running in real containers.

"If you get a Vera CPU rack, it's got over 20,000 cores available for agents to operate quickly and complete these tasks as fast as possible," DeLaere explained during the briefing. "Because Vera CPU was designed specifically for agents with its high bandwidth and its high single-threaded performance, we see that it really accelerates the agentic tool calling and execution of code … and having those separate is really how we can achieve these tremendous agentic workload benchmarks."

Nvidia has been preparing to take on the CPU dominance of AMD and Intel at a time when demand for agentic workloads has shaken up the processor market. Arm has even moved beyond its long-standing licensing model to launch physical chips.

Nvidia first announced plans for a standalone CPU business at the turn of the year, with Meta and CoreWeave among early customers. Amazon Web Services is also lined up to offer Vera, which it is billing as an additional compute option for agentic workloads.

Vera entered full production in the spring and is set to replace Grace as Nvidia's flagship processor. It also marks a departure from Nvidia's traditional reliance on Arm designs for CPUs, as it features a core design developed largely in-house, although it still uses the Arm v9.2-A architecture.

In addition to being available as a standalone processor and a dedicated rack platform, Vera will provide the CPU portion of the Vera Rubin Superchip, also scheduled for release in the second half of this year.