AMD announces general availability of its Zen 6-based 6th Gen EPYC server processors

AMD announced on 10 November 2026 that its 6th Gen EPYC SP7 server processors, using the company's Zen 6 design, are generally available. In a press release datelined Santa Clara, California, the company said server makers, both OEMs and ODMs, are already shipping systems based on the chips, and that Microsoft Azure and Oracle Cloud Infrastructure (OCI) will offer cloud instances soon.

AMD describes the family as a set of processors tuned to different jobs: agentic AI, enterprise, cloud, high-performance and general-purpose computing. Dan McNamara, who leads compute and enterprise AI at AMD, said AI is widening what the CPU does at each level of the data centre.

What AMD claims for performance

The performance figures are AMD's own, taken from its release and footnotes, and several are estimates. AMD says a dual-socket EPYC 9996 system scores up to 3.7 times an Nvidia Vera system on SPECrate integer, and up to 2.2 times a dual-socket Intel Xeon 6980P system. Its footnotes say the Vera score is an AMD internal estimate based on Phoronix's published testing.

AMD names three CPU roles in agentic AI: general services, hosts for GPU nodes, and the sandboxes where agents run. It says its high-frequency host-node parts, which reach 5.3 GHz, give up to 1.8 times the output throughput of the previous EPYC generation when running LLMs. Per the footnotes, that is a projection: AMD measured the earlier generation over PCIe Gen 5 and extrapolated to Gen 6, using a Qwen3-30B-A3B model and an Nvidia B200 GPU.

AMD also makes a rack-power claim. It estimates that EPYC 9996 servers can host about 1.18 times as many agents within a 35 kilowatt rack as servers left on default power-management settings. The footnote ties this to a combined processor and memory power limit 15 percent below default, with an estimated 1 percent average loss in per-server performance drawn from tests of a 5th Gen EPYC part. AMD adds that agent counts are estimates, that real capacity will vary, and that several of these figures are preliminary and subject to change.

Racks, customers and availability

AMD is also offering the processors in three Helion agentic AI rack designs. Helion Liquid is meant to sit beside its Helios rack-scale systems in large deployments. Helion Air is for adding agent capacity in existing air-cooled facilities. Helion Air Inference combines CPU execution with local GPU inference in a single rack. AMD says some configurations reach 73,728 threads per rack.

AMD says Meta is a lead customer, validating platforms based on the new chips ahead of deployment at scale. Azure's HDv2 instances are in preview, and Azure is also bringing HXv2 for electronic design automation and technical computing, plus ND MI455X v7, built on Helios, for AI inference. AMD says OCI's Compute E7 instances will use the processors.