AMD 发布面向智能体 AI 的 CPU 中心机架设计 Helion
AMD Unveils CPU-Centric Helion AI Rack for Agentic AI < Semiconductor < 기사본문 - The Elec Inc.
AMD 在 OCP Global Summit 上发布 Helion，一套以 CPU 为中心的 AI 数据中心计算基础设施设计，用于同时运行大量 AI 智能体。与堆叠大量 GPU 的 Helios 不同，Helion 面向智能体反复执行程序、访问数据库带来的 CPU 需求；面向大型数据中心运营商的配置可单机架容纳最多 32,768 个基于第六代 EPYC Venice 的 CPU 核心，模型推理请求则交给独立 GPU 系统。企业配置将 EPYC CPU 与 Instinct MI350P GPU 放在同一机架，最多容纳 16 块 GPU，并集成 token router 把请求分流到高性能模型或自托管开源模型以降低 AI 运营成本。AMD 将以开放参考设计形式提供 Helion，向 OCP 贡献规范，Cisco、Dell、HPE、Lenovo、Supermicro 等八家服务器厂商正在开发相关系统，计划 11 月上市。
推荐理由
AI 评分 58
AMD 把 AI 数据中心设计从 GPU 转向 CPU，并给出面向智能体负载的机架配置与开放参考设计路径。
主题
正文
当前显示原文，中文全文译文尚未就绪。
JEONG IL JOO; Published 2026.10.13 14:29; Reply 0
AMD announced on Oct. 12 at the Open Compute Project (OCP) Global Summit in San Jose, California, that it has unveiled Helion, a computing infrastructure design for artificial intelligence (AI) data centers built around central processing units (CPUs).
Unlike its existing Helios platform, which combines large numbers of graphics processing units (GPUs), Helion is designed to run large numbers of AI agents simultaneously.
AMD said the spread of AI agents is expanding computing demand in AI data centers beyond GPUs to CPUs. AI agents do more than call large language models (LLMs); they repeatedly execute programs and access databases. These tasks require substantial CPU resources, the company said.
Helion configurations can vary according to customer requirements. A configuration designed for large data center operators uses CPUs alone, without GPUs. It can accommodate up to 32,768 cores based on AMD’s sixth-generation EPYC Venice server CPU in a single rack.
In this configuration, Helion handles AI agent execution, while requests for AI model inference are sent to separately installed GPU systems. This allows Helion to work alongside large GPU racks such as Helios, with each system serving a distinct role.
An enterprise configuration combines CPUs and GPUs in a single rack. It pairs AMD EPYC CPUs with Instinct MI350P GPUs to handle both AI agent execution and model inference within one system. The rack can accommodate up to 16 GPUs.
Helion also integrates a token router, which directs requests from AI agents to either the latest high-performance models or open-source models. AMD said that routing some requests to open-source models running on a company’s own servers, rather than sending every request to expensive external high-performance models, can reduce AI operating costs.
AMD will offer Helion as an open reference design rather than a finished server product. Based on OCP standards, the design allows customers to adjust CPU core counts and power consumption to their needs. It supports both air and liquid cooling, enabling installation in existing data centers. Eight server makers, including Cisco, Dell Technologies, HPE, Lenovo, and Supermicro, are developing related systems.
AMD said it would contribute Helion specifications to the OCP. Systems developed by server vendors are scheduled to become available in November.