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 introduced a reference rack design called Helion for agentic AI servers, one that can hold as many as 32,768 Zen 6 "Venice" cores in a CPU-only configuration. The design, based on the Open Compute Project's ORv4 specification, was detailed in a briefing with press ahead of the OCP Global Summit. AMD will not sell the racks itself. Instead, it is contributing the design to OCP, and OEMs and ODMs will take it to market.

Helion is standardized around an ORv4 rack, which is slightly wider and taller than a standard IT rack. It is built to support 44 OpenU trays measuring 21mm wide and 48mm tall, slightly larger than a standard EA1-310 tray. AMD says the reference design can be deployed with air cooling or direct liquid cooling, and it complies with OCP's Data Center Modular Hardware System standard, which OCP describes as bringing interoperability between key elements of data center, edge, and enterprise infrastructure.

In a CPU-only deployment, Helion can hold up to 128 EPYC 9996 CPUs, producing the 32,768-core figure per rack. That configuration resembles an agentic AI sandbox system connected to an inference system, a model AMD points to as being in use today in Meta's Muse and OpenAI Dots. In that model, each user gets a sandbox with a slice of compute from a host system backed by an inference server, with some workloads such as building software handled on CPU-only servers while heavier AI inference runs elsewhere.

AMD positions Helion as more flexible than Helios, which is built for AI training and inference, with Helion focused primarily on CPUs. The design can alternatively support up to 16 MI350P GPUs and 102 Tbps of networking, though adding GPUs reduces CPU capacity. AMD says that hybrid configuration is aimed at enterprises, offering an all-in-one rack for data management, networking, and AI usage.

AMD did not name partners supporting Helion during the press pre-briefing. The company says CEO Lisa Su will announce partners during the lead keynote at the OCP Global Summit later today. AMD also did not share deployment partners, though it suggested that Helios customers will also deploy Helion racks.

AMD says the Helion rack will be available starting in the first half of 2027. That timing would likely coincide with Zen 6 deployments, which are expected to begin ramping up later this year. For US hyperscalers and enterprises building out CPU capacity around agentic AI, the reference design offers a standardized path that hardware vendors, not AMD, would bring to market.

More AI news from TechManNews.