Nvidia and AMD split over CPU metrics for AI agents

Nvidia and AMD are framing a new CPU performance debate around AI agent workloads, where procurement teams may choose between faster sequential execution and higher concurrent capacity. Nvidia is promoting Vera around what Bank of America analyst Vivek Arya described as “maximum single-threaded performance at scale,” arguing that agent workflows depend on low latency as CPUs and GPUs repeatedly coordinate tool calls, code execution, data retrieval and orchestration.

Vera uses 88 custom Olympus ARM architecture cores, 1.2TB/s memory bandwidth and 3.4TB/s on-chip interconnect bandwidth. Nvidia says this design helps reduce GPU wait time and improve utilization, with a monolithic compute die intended to support coherency and limit cross-die latency. AMD is countering with an EPYC strategy focused on rack throughput, service density and fixed-power efficiency.

AMD says EPYC 9965 (Turin) delivers approximately 2.4x the rack-level throughput of Vera’s baseline in a 100kW rack deployment, while EPYC 6 (Venice) is projected to reach 3.3x. Bank of America expects AMD’s AI 2026 Technology Day to sharpen that response as the server CPU market could expand to $170 billion by 2030. The broader contest is over which KPI data center buyers adopt for future CPU purchases.