32,768 cores and 74 petaflops: AMD’s monster new rack for AI agents could be too much for traditional data centers

AMD's proposed rack combines 32,768 CPU cores with 16 AI accelerators Traditional data centers may need new power and cooling infrastructure AMD divides agent workloads into personal, enterprise, and technical computing categories AMD is developing a rack-scale computing system for artificial intelligence agents, with configurations designed for personal, enterprise, and technical workloads.

Dubbed Helion Agentic AI Rack, it supports up to 32,768 Zen 6 x86 CPU cores, 16 Instinct MI350P accelerators, and 74 petaflops of FP4 computing performance.

The proposed platform also supports 102 Tbps of networking bandwidth and uses modular hardware designed for different agent computing requirements.

A rack built around three kinds of AI agents AMD groups agent workloads into three categories: personal assistants, enterprise applications, and technical computing, each with different processing requirements.

Personal agents need dense, energy-efficient servers; enterprise applications require a balance between agent capacity and performance; while technical workloads prioritize processing speed.

AMD allows operators to customize CPU core density, processing performance, and power consumption according to their workloads.

The hyperscale configuration uses CPU-only infrastructure, while the enterprise version combines processors and graphics accelerators within an integrated rack.

For personal agents, the system uses 256-core Venice processors operating at 400W, prioritizing processing density and power efficiency.

Enterprise workloads use the 256-core Venice processors rated at 600W, while technical computing uses 96-core, high-frequency Venice processors.

The enterprise configuration separates agent execution, inference, and general-purpose computing into distinct server nodes.

CPU sandbox nodes handle agent execution, Instinct MI350P accelerators support inference, and general-purpose servers accommodate databases, storage, and technical computing.

AMD also includes an integrated token router for directing agent model requests across frontier and open-source models.

The rack supports programmable networking, an open-source software ecosystem, and ORv3 44OU hardware compliant with Open Compute Project specifications.

Why traditional data centers may need substantial changes Traditional data centers are built for standard cloud or legacy enterprise computing, not racks packing up to 32,768 Zen 6 CPU cores alongside heavy GPU configurations.