NVIDIA Vera CPU Outperforms Intel Sapphire Rapids by 2.2x in Agentic AI Benchmark
Cloud AI inference platform DeepInfra released benchmark results showing that NVIDIA's new Vera CPU is 2.2 times faster than Intel's Sapphire Rapids in containerized Agentic AI workloads. In the tests, the Vera CPU achieved a runc latency of 29ms compared to Sapphire Rapids' 64ms. As Agentic AI workloads become more prevalent, optimizing CPU performance for containerized environments is crucial for reducing latency and operational costs in AI infrastructure. This benchmark highlights NVIDIA's growing competitiveness in the data center CPU market, challenging traditional x86 giants like Intel and AMD. The benchmark was conducted under uniform conditions using 20 physical cores, 1 thread per core, a single NUMA node, and CPU confinement. NVIDIA Vera also outperformed AMD's Zen5 Turin and Intel's Granite Rapids, achieving the lowest latency in both runc and Kata container deployments.
## BACKGROUND
Agentic AI refers to autonomous AI systems that can pursue goals, use tools, and take actions within defined constraints. To run these workloads efficiently, cloud providers often deploy them in lightweight containers using runtimes like runc, which relies on Linux kernel features to isolate processes. NVIDIA's Vera CPU is specifically designed to handle these agentic workloads, featuring 88 Olympus cores and high-bandwidth LPDDR5X memory.