NVIDIA just officially said Vera is now shipping at scale, with AWS receiving its 1st CPU server and Vera Rubin GPU.

This chip targets CPU work around agents, including Python execution, tool calls, retrieval, orchestration and sandboxed code.

Those jobs can become a bottleneck because one agent may repeatedly run code and coordinate tools around each GPU inference step.

NVIDIA built Vera with 88 custom Olympus cores and up to 1.2TB/s memory bandwidth, claiming up to 1.8x faster performance on selected agentic workloads than x86 CPUs.

Independent Phoronix testing found Vera 10% ahead of AMD EPYC 9575F on the geometric mean across permitted workloads, though NVIDIA restricted the test suite to Vera-relevant workloads.