Helion Racks

Alongside the silicon, AMD introduced Helion, a rack-scale reference design. The argument behind it is that agentic AI has created a new tier of infrastructure that does not map onto existing racks.

AMD splits the agentic pipeline across three infrastructure tiers:

• GPU servers, the host node—high-frequency cores and maximum CPU-to-GPU bandwidth
• Agent CPU servers, the sandboxes—the maximum number of performant cores per watt
• General purpose CPU servers—a balance of performance, core count, power and cost

The middle tier is the new one. Agent sandboxes are short-lived, isolated environments spun up and torn down by the thousands, and they are pure CPU work. Three Helion configurations address it:



A fully populated Helion Liquid rack reaches 73,728 threads at around 120 kW. AMD compared that against competing racks in the same 44/48 OU ORv3 footprint: Intel "Clearwater Forest" at 288 cores reaches 36,864 threads, NVIDIA Vera "Olympus" at 88 cores reaches 45,056 threads in 48 U MGX, and an Arm Neoverse design at 136 cores reaches 20,672.

The software side is the ROCm Agentic Toolkit, a reference stack covering the agent harness, scheduling and placement, session management, sandboxing, caching and memory, and a policy-based model router. AMD emphasized standard racks, standard servers and open-source tooling throughout. More than ten "Helion Certified" racks are planned, with the first available starting November 2026.