AMD's 6th Gen EPYC server CPUs ship on Zen 6, with a dual-socket 9996 claimed up to 2.2x a Xeon 6980P

AMD's 6th Gen EPYC server SP7 processors are generally available on the new Zen 6 architecture, spanning agentic AI, enterprise, cloud, HPC and general-purpose workloads, and are shipping from leading server OEMs and ODMs. AMD says five prior EPYC generations repeatedly raised server CPU performance, energy efficiency and platform choice, with 5th Gen EPYC the fastest-ramping and highest-volume generation in its history. Microsoft Azure and OCI will offer cloud instances on the processors: Azure HDv2 is in preview, Azure HXv2 and the AMD Helios-based Azure ND MI455X v7 are listed, and OCI Compute E7 will be powered by 6th Gen EPYC; Meta is a lead customer validating 6th Gen EPYC platforms ahead of deployment at scale.

AMD calls the EPYC 9996 the world's best general-purpose processor, leading on performance, performance per CPU watt, performance per CPU dollar and performance per socket. AMD's stated comparisons are against Intel Xeon 6980P and Nvidia Vera: a dual-socket EPYC 9996 system delivers up to 3.7x the performance of a Vera-based system and up to 2.2x the performance and 2.0x the performance per dollar of a Xeon 6980P system. The comparison uses SPECrate 2017_int_base with scores published on spec.org as of November 5, 2026, while the Vera and Arm AGI figures are AMD internal estimates as of October 10, 2026; the 2P EPYC 9996 configuration is 512 cores at 600 W default CPU power, $14,904 at 1Ku, scoring 5470.

Agentic AI is split into three CPU roles, general-purpose services, GPU host nodes and agent sandboxes, and 6th Gen EPYC offers a workload-optimized part for each. AMD says high-frequency host-node processors at up to 5.3 GHz deliver up to 1.8x the LLM output throughput of previous-generation EPYC, and high-density parts provide up to about 3.3x agents per CPU watt over Nvidia Vera and about 2.5x over Intel Xeon 6980P for hosting agent sandboxes. The 1.8x throughput comes from AMD internal testing of a 96-core high-frequency host node against 5th Gen EPYC 9575F, measured on a PCIe Gen 5 x16 host link and projected to PCIe Gen 6 x16 from host-to-GPU link bandwidth, using Qwen3-30B-A3B, vLLM CPU-offload and one NVIDIA B200; agent counts are estimated using CPU thread resources as a proxy, with actual capacity varying by workload, model, memory and orchestration.

The report covers AMD's announced portfolio, availability status and comparison data: systems are shipping from leading server vendors, Azure HDv2 is in preview, Azure HXv2, ND MI455X v7 and OCI Compute E7 are listed as forthcoming, and Meta is still validating platforms. The intended reader is an engineer sizing server CPUs and agent infrastructure capacity; AMD states no follow-up validation plan in the text.

AMD states these are preliminary performance estimates based on engineering projections or measurements as of specific dates and subject to change; the Vera and Arm AGI scores are internal estimates, and agent counts are derived using CPU thread resources as a proxy, with actual capacity and throughput varying by workload, model, memory, software, orchestration and system configuration. The company also says "best" reflects the strongest overall combination across evaluated criteria for general-purpose server CPUs and does not imply leadership in every individual workload or metric.