AMD Unveils Helion Rack Design for Agentic AI Workloads

AMD introduced Helion on Oct. 12 as an open rack-scale reference design for agentic artificial intelligence workloads, separating CPU-only systems for hyperscalers from mixed CPU-GPU configurations for enterprises.

The CPU-only configuration supports up to 32,768 cores using sixth-generation AMD EPYC processors. The enterprise configuration supports as many as 16 Instinct MI350P GPUs and 102 terabits per second of networking bandwidth.

Agentic AI systems can carry out multistep tasks, use tools and route work among different models. Helion is designed for that range of workloads rather than relying on a single large GPU cluster.

“Agentic AI is not a single workload. It spans personal agents, enterprise applications, code execution and complex workflows, each with different compute requirements,” Dan McNamara, AMD’s senior vice president and general manager of Compute and Enterprise AI, said.

The CPU-only configuration targets hyperscale operators with separate accelerated inference infrastructure. The enterprise configuration combines EPYC CPU nodes, MI350P GPU nodes and general-purpose computing in one rack.

Customers can mix CPU blade types and add GPU blades based on workload requirements. The enterprise design uses two independent five-rack-unit GPU servers, each containing eight MI350P GPUs. The two GPU servers do not use a scale-up GPU fabric between them.

Helion supports liquid-cooled and air-cooled versions and is built around a 21-inch ORv3, 44OU rack. The design includes an integrated token router that directs model calls between frontier and open-source models.

Helion extends AMD’s rack-scale systems strategy but serves a different purpose from AMD’s Helios platform, which targets frontier-scale training and large-scale inference. Helion is designed to keep agentic inference within a rack.

AMD plans to contribute the Helion specification to the Open Compute Project. Configurations developed with ecosystem partners, including Dell, Hewlett Packard Enterprise (HPE), Lenovo, QCT, Supermicro and Wiwynn, are planned to be available in November 2026.

“With AMD Helion, we are bringing AMD EPYC performance, open software and flexible rack designs together so our partners can build systems around customer needs and make efficient use of available power,” McNamara said.