AMD has unveiled Helion, a rack-scale computing architecture created to support agentic AI workloads, offering enterprises a flexible combination of CPUs and GPUs.
Announced at the Open Compute Project (OCP) Global Summit, Helion reflects a growing emphasis on CPUs in AI infrastructure. While GPUs have dominated investment in AI hardware, the explosive growth of AI agents is driving demand for CPUs, which are well-suited for tasks like executing code and coordinating complex workflows.
AMD’s Helion architecture offers two primary configurations. The first is a CPU-only rack equipped with sixth-generation AMD EPYC processors, supporting up to 32,768 CPU cores. Built for hyperscale data centers, this configuration handles agent execution while connecting to separate GPU systems for AI inference.
The second configuration targets enterprises that want a more integrated solution. It combines EPYC processors with up to 16 AMD Instinct MI350P GPUs, providing 74 petaflops of FP4 compute performance. AMD lists networking bandwidth of up to 102 terabits per second for Helion, although it hasn’t published real world deployment results showing that capacity.
AMD says the new platform improves the efficiency of AI infrastructure by allowing customers to allocate computing resources based on the requirements of individual applications.
“Every rack-scale AI product of the past three years placed the GPU at the center,” said Brendan Burke, Research Director at Futurum Group. “Helion is the first rack a major silicon vendor has sold for AI in which the GPU is a configurable option, and in the hyperscale option absent entirely. CPU:GPU rebalancing now has a procurable unit with agentic capacity in cores and watts rather than accelerators.”
Multiple AI Models
To serve enterprises that use multiple AI models, the platform includes an integrated token router that directs AI requests between different models, including proprietary frontier models and open-source alternatives. So organizations can route AI processing across computing resources rather than relying on a single model or infrastructure environment.
This shift could expand demand for server processors as enterprises deploy more autonomous AI applications. AMD has forecast that the server CPU market could exceed $200 billion by 2030, compared with an earlier projection of approximately $60 billion.
Yet separating CPU-based agent execution from GPU-based inference could present challenges. Such a system could create more network traffic, so companies might need extra networking infrastructure to prevent delays as AI agents repeatedly exchange data with inference systems. It’s possible that these costs could reduce the savings achieved through more efficient use of GPUs.
Software compatibility presents another challenge. AMD must demonstrate that its ROCm software ecosystem and Helion’s token-routing capabilities can support production workloads with the reliability enterprises require.
Helion also gives AMD another way to compete with NVIDIA, whose AI infrastructure business has benefited from demand for tightly integrated computing systems. NVIDIA is developing its own CPU-based rack architecture for agentic AI and has contributed Vera Rubin hardware designs to OCP.
AMD is pursuing an open architecture that allows hardware partners to customize systems for individual customers. The company plans to contribute Helion’s specifications to OCP, supporting broader industry adoption.
Cisco, Dell Technologies, HPE, Lenovo and Supermicro are among the vendors supporting the initiative. Dell is extending its PowerRack portfolio with a Helion-based reference design.
AMD expects partner-developed Helion configurations to become available in November.
