Nvidia Vera System Leads Dual-Socket AMD EPYC 9755 System by 4% in PyBench

A two-socket server built around Nvidia’s Vera CPUs scored 4% higher on average than a dual-socket AMD EPYC 9755 server in PyBench, giving the Vera system an edge on Python workloads in a comparison published Oct. 30.

The test measured Python performance, not overall server performance. Its result describes the two server configurations as tested rather than a direct comparison of individual processors.

The Vera system combined two processors for 176 cores and 352 threads. The AMD system paired two EPYC 9755 chips for 256 cores and 512 threads. Each server had 1.5 terabytes of memory, but the systems used different memory types and speeds.

A separate SPEC CPU 2026 integer-throughput comparison put the Vera system 1.5% ahead of the EPYC 9755 system. That figure was an estimate and was not submitted to SPEC. Nvidia had previously estimated a 3% lead in a different comparison, so the two percentages describe separate measurements.

Nvidia designed Vera for data-center work involving agentic artificial intelligence and reinforcement learning. Each CPU has 88 custom Olympus cores and LPDDR5X memory bandwidth of up to 1.2 terabytes per second. In dual-socket systems, NVLink-C2C can provide up to 1.8 terabytes per second of bandwidth between the CPUs.

Python performance is relevant to agentic software, but PyBench does not show how quickly a system completes a full AI-agent task. No standard test yet measures how quickly CPUs complete full agentic tasks.