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, dubbed Helion Agentic AI Rack, boasts an impressive 32,768 CPU cores, 16 AI accelerators, and 74 petaflops of computing power. This system is tailored to handle various AI agent workloads, including personal assistants, enterprise applications, and technical computing. The rack supports different configurations for each type of workload, from energy-efficient servers for personal agents to high-frequency systems for technical computing.

AMD categorizes agent workloads into three types: personal, enterprise, and technical computing. Personal agents require dense, power-efficient servers, while enterprise applications demand a balance between capacity and performance. Technical workloads prioritize processing speed above all else. The rack can be customized to meet these varying requirements, allowing operators to adjust CPU core density, processing performance, and power consumption as needed.

Traditional data centers, designed for standard cloud or legacy enterprise computing, may struggle to accommodate such a high-density rack. These facilities may lack the necessary power delivery, cooling capacity, and network infrastructure to support a system with up to 32,768 CPU cores and 16 AI accelerators. Additionally, the diversity of workloads, ranging from personal agents to technical computing, necessitates infrastructure capable of accommodating different server types and operating requirements.

The proposed rack also presents challenges in terms of networking and infrastructure compatibility. With the ability to support up to 102 Tbps of bandwidth and an integrated programmable token router, the rack demands networking equipment and software capable of handling such high-speed data transmission. Existing data centers may need to upgrade their networking infrastructure and ensure compatibility with the rack's programmable networking capabilities.

AMD's Helion Agentic AI Rack is set to launch in November 2026. However, operators will need to carefully consider the costs and cooling requirements associated with upgrading their facilities to support this cutting-edge AI computing platform.