Single-socket vs dual-socket server

A single-socket (1P) server has one CPU, and a dual-socket (2P) server has two CPUs joined by a cache-coherent link. Each CPU in a 2P server owns its memory channels and PCIe devices. A core that reaches memory or a GPU on the other socket waits longer than it would for local resources. That delay is the non-uniform memory access (NUMA) penalty.

On AMD EPYC 9005, a second socket doubles memory channels from 12 to 24, but not lanes. The sockets connect over Infinity Fabric (xGMI) links that each take up to 16 of a socket's 128 PCIe 5.0 lanes. A 2P EPYC server therefore exposes 128 to 160 lanes. Intel's Xeon 6 joins sockets over Ultra Path Interconnect (UPI) 2.0 at up to 24 GT/s, and two sockets offer 192 PCIe 5.0 lanes. NVIDIA's DGX B200 pairs eight Blackwell GPUs with two Xeon Platinum 8570 CPUs, 112 cores in total. AMD sells P-suffix EPYC parts for single-socket servers only.

Choose 1P when one socket's cores, memory channels and lanes cover the build, which also avoids cross-socket traffic. Choose 2P when the build runs out of cores or memory capacity. On EPYC, a second socket adds at most 32 lanes, so it does not fix a lane shortage.

Sources

Last verified October 7, 2026.