AMD debuts Helion reference design for agentic AI servers with up to 32,768 Zen 6 ‘Venice’ cores — rack design ships next year in CPU-only and all-in-one configurations

AMD has unveiled a new reference rack-scale design called Helion for the data center industry. This standardized 21-inch rack design is part of the Helion family and aims to complement the previously announced Helios rack. The Helion design can accommodate up to 32,768 Zen 6 cores (codenamed 'Venice') in a CPU-only setup, leveraging the dense 256-core EPYC 'Venice' 9996 processor.

The rack design adheres to the Open Compute Project's ORv4 specification, which is slightly wider and taller than a standard IT rack. It supports 44 OpenU trays, measuring 21mm by 48mm, making them slightly larger than a standard EA1-310 tray. Helion is designed for compatibility with both air-cooled and direct liquid-cooled deployments, and it meets the OCP's DC-MHS (Data Center Modular Hardware System) standard.

AMD is contributing this reference design to OCP and expects OEMs and ODMs to bring the design to market. Although AMD did not reveal specific partners for Helion, the company confirmed that CEO Lisa Su will announce partners at the OCP Global Summit. Helion differs from Helios, which is designed for AI training and inference, as Helion is a more flexible CPU-focused rack solution.

In a CPU-only configuration, the rack can hold up to 128 EPYC 9996 CPUs, enabling a potential 32,768 core per rack setup, which could be used for agentic AI sandbox systems.

The design also supports up to 16 MI350P GPUs and offers 102 Tbps of networking, though this hybrid version prioritizes data management, networking, and AI usage over CPU capacity. AMD plans to release the Helion rack in the first half of 2027, aligning with the expected ramp-up of Zen 6 deployments later in the year.