Eigen RadarAI
Archive
Analysis

AMD’s Helion puts agent execution and model inference in flexible racks

AMD has announced Helion, an open rack design for AI agents. It supports processor-only systems connected to separate model inference infrastructure, as well as enterprise racks combining processors and graphics accelerators. Air and liquid cooling are both supported. Partners are developing configurations, while the proposed contribution to the Open Compute Project concerns a shared infrastructure blueprint.

Artificial Intelligence··Morning
Two distinct processor racks connected by a supported cable in one datacenter aisle.

Helion separates agent work from model inference

AMD announced Helion on October 12 as an open rack blueprint for AI agents, software that calls models and executes tasks in several steps. The design follows specifications from the Open Compute Project, an industry initiative for shared hardware designs. It extends AMD’s open-platform approach with Helios to computing jobs whose requirements vary between agent execution and model inference. Partners can assemble systems around the work customers need to run.[1], [2]

In the processor-only hyperscale arrangement, EPYC processors provide dedicated capacity for executing agent tasks. Model inference runs on separate infrastructure to which the rack connects. The enterprise arrangement brings EPYC processor nodes, Instinct MI350P graphics accelerators and general-purpose computing together. Here, agent execution, inference and supporting workloads can reside within one rack. The choice changes where the work runs rather than requiring every customer to adopt the same combination of hardware.[1], [2]

Cooling and compute adapt to different workloads

Helion supports both air cooling and liquid cooling. AMD describes choices in processor performance, core density, power and acceleration that partners can match to customer requirements. Personal agents, enterprise applications, code execution and complex workflows are among the jobs the company identifies. The integrated design also includes general-purpose components for database, storage and technical computing, alongside the components assigned to agent execution and graphics-processor inference.[1], [2]

AMD also describes an integrated router intended to distribute agents’ model calls between frontier models and open models. Its role concerns selecting where a model request goes within the described system. The rack blueprint addresses execution resources and connections across different kinds of computing work; the announcement presents the router as part of that planned infrastructure.[1]

Partners are developing the open blueprint

AMD plans to contribute the Helion specification to the Open Compute Project so the community can develop and adopt it. Configurations remain under development with ecosystem partners. Cisco, Dell, HPE, Lenovo and Supermicro are among the companies named in the announcement. Their support concerns building and integrating systems, including cooling, management and adaptation to existing infrastructure. The launch establishes the shared design and partner effort, with deployment work still ahead.[1], [2]

References

  1. News sourceAMDAMD announces the Helion open rack design for agent workloads↩1↩2↩3↩4↩5
  2. News sourceThe RegisterHelion racks combine agent execution and inference options↩1↩2↩3↩4
AI safety and regulationChip wars and the supply chainThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →