AMD says Helion CPU-only rack design supports up to 32,768 cores

Tom’s Hardware reports that the maximum configuration corresponds to as many as 128 EPYC 9996 processors, each with 256 cores. AMD’s announcement names 6th Gen EPYC processors but does not specify that model. The publication also describes Helion as based on the Open Compute Project’s ORv4 specification, with a 21-inch-wide rack designed for 44 OpenU trays, each 48 millimeters tall.

AMD says the blueprint supports air- or liquid-cooled configurations. The announcement and coverage do not provide measurements of power use or cooling performance, so those design options do not establish how a finished system would perform in operation.

The proposed division of work puts agent execution on the CPU-only rack and model inference on separate infrastructure. That means the CPU-only configuration is not presented as a standalone rack for both functions. The available reports do not include independent comparisons showing how this arrangement performs on particular agent workloads relative to other systems.

How does the mixed Helion design differ?

AMD also outlined an all-in-one configuration aimed at enterprise agentic AI deployments. It combines EPYC CPU nodes, Instinct MI350P GPU nodes and general-purpose compute in one rack, rather than placing agent-execution CPUs apart from inference infrastructure.

Tom’s Hardware reports that the mixed configuration can include up to 16 MI350P GPUs and 102 Tbps of networking. Those are reported design maximums for the mixed system, not specifications for the CPU-only rack.

AMD’s rationale is that agentic AI can involve workloads with different computing needs, and that partners should be able to adjust the mix of CPUs, GPUs and rack components. That is the company’s design goal, not independent evidence that either layout performs better for a particular workload.

When could Helion systems reach customers?

Helion is a reference design for systems partners to build from, not a finished rack AMD says it will sell itself.

AMD’s November 2026 timing is a plan for partner configurations, not confirmation that systems will ship then or be generally available. Which versions partners will offer, and when customers can obtain them, remain unresolved.

The announcement lays out two approaches for infrastructure buyers: dedicate a CPU-only rack to agent execution and connect it to separate inference systems, or combine CPUs, GPUs and general-purpose compute in one rack. The published reports provide no independent workload benchmarks or measured comparisons for throughput, power use, operating cost or performance per watt. Those results will be needed to show where either design fits among AI computing options.