Photo: Genevieve Martin, OLCF at ORNL, CC BY 2.0 (https://creativecommons.org/licenses/by/2.0), via Wikimedia Commons (source)
AMD on Monday, Oct. 12, rolled out Helion, a rack-level reference design for running AI agents, tied to an Open Compute Project layout and timed for the OCP Global Summit in San Jose, the company said in a GlobeNewswire statement.
The design sits alongside AMD's earlier Helios rackscale platform. Company materials say Helion is meant so buyers can balance how many CPU cores they deploy, how much power they draw and whether they also attach GPU acceleration for the work their agents do. Dan McNamara, who leads Compute and Enterprise AI at AMD, said agents are not one kind of job: they cover personal assistants, business software, code runners and longer workflows, each with different hardware needs.
Two example layouts
In one layout described by AMD, a rack filled only with 6th Gen EPYC processors can reach as many as 32,768 cores. That version is pitched at hyperscale operators that want a dense place to run agents and then talk to separate systems that serve the underlying models. A second layout mixes EPYC nodes with Instinct MI350P GPU nodes and ordinary general-purpose servers on the same rack for enterprise buyers that want one package.
DatacenterDynamics said the dense CPU version uses EPYC chips known as "Venice" and that AMD is offering both liquid-cooled and air-cooled options. The company said partners can change the CPU-to-GPU mix to hit targets for speed, watts and cost.
Open contribution and partners
AMD said it intends to hand the Helion specification to OCP so others can extend it. Systems built with partners are still in development and are planned to ship in November 2026, according to the company release.
Supporting statements in the release came from Cisco, Dell Technologies, Hon Hai Technology Group (Foxconn), HPE, Lenovo, QCT, Super Micro and Wiwynn. DatacenterDynamics named Cisco, Dell, Hon Hai, HPE, Lenovo, Super Micro and Wiwynn among firms working toward the launch.
McNamara said the goal is to let partners assemble systems around what each customer needs and use available power efficiently, rather than forcing a single fixed mix of chips.
What's next
Partner-built Helion racks remain in development ahead of the November 2026 availability target AMD stated in its release. DatacenterDynamics likewise reported that configurations are still being finished with ecosystem partners before that planned window.
Sources
- Advanced Micro Devices (GlobeNewswire): Introducing AMD Helion, Agentic AI Compute at Rack Scale (2026-10-12)
- DatacenterDynamics: AMD unveils CPU-only Helion rack to support agentic AI workloads (2026-10-12)
Motion Media News corrects errors promptly. To report one, email hello@motionnews.studio.
