AMD Unveils Helion Rack Architecture for AI Agents With Up to 32,768 Zen 6 Cores

Advanced Micro Devices is developing a rack-scale computing architecture tailored for artificial intelligence agent workloads across personal, enterprise, and technical computing environments. Designated the Helion Agentic AI Rack, the system supports up to 32,768 Zen 6 x86 CPU cores, 16 Instinct MI350P accelerators, and 74 petaflops of FP4 computing performance, as detailed in a report by TechRadar Pro .

The architecture supports up to 102 Tbps of networking bandwidth and uses modular hardware designed to address three distinct classes of agent operations. Personal assistants require dense, power-efficient configurations; enterprise workloads balance agent volume with throughput; and technical computing prioritizes raw execution speed. Operators can adjust core density, compute performance, and power draw according to specific workload demands.

Hardware configurations vary by use case. For personal assistant deployments, the rack uses 256-core Venice processors running at a 400W power envelope for maximum density and efficiency. Enterprise workloads employ 256-core Venice processors rated at 600W, while technical computing tasks rely on 96-core, high-frequency Venice chips. AMD plans a CPU-only configuration for hyperscale operators alongside an integrated CPU-GPU rack for enterprise deployments.

Within the enterprise configuration, processing tasks are partitioned across specialized nodes. Dedicated CPU sandbox nodes run agent execution environments, Instinct MI350P accelerators handle neural network inference, and general-purpose servers support underlying databases, storage, and technical computing tasks. The rack also integrates a programmable token router designed to distribute model queries across frontier and open-source AI models.

The system complies with Open Compute Project specifications under the ORv3 44OU standard and supports programmable networking and an open-source software ecosystem. However, the high-density layout presents structural challenges for conventional data center facilities built for standard cloud or legacy enterprise hardware. Combining up to 32,768 Zen 6 cores with GPU accelerators requires power delivery, cooling capacity, and high-throughput networking that older facilities may lack.

The mix of dense 256-core processors and high-frequency 96-core units also complicates facility planning, forcing operators to allocate power and cooling non-uniformly across different server nodes rather than standardizing on identical racks. AMD has scheduled the Helion Agentic AI Rack for availability in November 2026, though the company has not yet published cooling specifications or facility upgrade cost projections.