AMD Helion rack separates agent execution from GPU inference
The rack blueprint gives partners a choice between CPU-only agent capacity and mixed CPU-GPU inference, with builds scheduled for November 2026.
| AMD Helion | |
|---|---|
| CPU cores | Up to 32,768 Zen 6 x86 CPU cores |
| GPUs | Up to 16 AMD Instinct MI350P GPUs |
| Networking bandwidth | Up to 102 Tbps of networking bandwidth |
| Rack | 21-inch ORv3 44OU DLC and air-cooled rack |
| CPU blade options | Dense 256 cores at 400 W; balanced 256 cores at 600 W; performance 96 cores at 5 |
| GPU inference blade | Eight AMD Instinct MI350P GPUs in a 5U server |
| General-purpose blade CPUs | AMD EPYC 9006 SP7, 9006 SP8, and 9006X SP7 |
| Ship date | Partner-built Helion configurations scheduled for November 2026 |
The reference design covers two layouts, a CPU-only hyperscale rack for agent execution and an enterprise rack that adds GPU inference blades.
The CPU-only layout can scale to 32,768 Zen 6 cores, while the enterprise layout adds up to 16 MI350P GPUs and up to 102 Tbps of networking bandwidth.
Blade choices range from dense 400 W EPYC 9006 chips to 96-core 5 GHz performance parts and balanced 256-core 600 W processors.
General-purpose blades cover database, storage, and technical compute workloads, and AMD says it will contribute the Helion specifications to OCP.
Cisco, Dell, HPE, Foxconn, Lenovo, and Supermicro are listed as partners, with builds scheduled for November 2026.
The page does not state rack price, total power draw, actual agents per watt, memory capacity, storage capacity, exact launch date, or exact SKU.
Die Brief reads the rack as a power-budget constraint, not a training fabric.
After amd.com. We did not report this. The pictures, if any, are theirs.
