AMD has unveiled the Helion rack-scale AI compute reference design, specifically engineered for AI agents. The architecture focuses on CPU-centric performance, allowing for flexible integration of GPUs, networking, and storage resources to handle complex workflows.
The flagship configuration supports up to 32,768 Zen 6 CPU cores, utilizing 128 units of 256-core Venice-based EPYC processors. This design is intended to provide dedicated execution environments for large-scale AI agent deployments, separating general-purpose compute from model inference.
For hybrid workloads, the system can integrate AMD Instinct MI350P GPU accelerators. A standard GPU-focused configuration supports up to 16 accelerators and offers a total network bandwidth of 102 Tbps, enabling simultaneous model inference and data processing.
AMD is contributing the Helion specifications to the Open Compute Project (OCP). Major hardware partners, including Dell, HPE, Lenovo, Supermicro, and Foxconn, are participating in the development, with commercial systems expected to launch in November 2026.
The platform offers diverse node types to balance power and performance, including high-density 400W nodes, balanced 600W nodes, and high-frequency 96-core nodes reaching 5GHz. The design supports both air and liquid cooling to manage thermal loads in high-density data centers.
