AMD today unveiled Helion, an open rack reference design built for agentic AI and evolved from the company’s existing rack-scale Helios AI platform. Helion comes in two distinct variants: a CPU-only rack built for hyperscalers, packing up to 32,768 Zen 6 cores powered by EPYC 9006-series Venice processors; and an all-in-one enterprise rack that combines EPYC CPU blades with as many as 16 Instinct MI350P GPUs. AMD intends to submit the Helion specification to the Open Compute Project.

Configurable Architectures for Dense Agent Execution and Inference

Both variants adopt a 21-inch ORv3 44OU rack built around OCP-compliant modular compute boards. They support either direct liquid cooling or air cooling, deliver up to 102Tbps of programmable fabric networking, and ship with an open-source software stack. The hyperscale configuration is CPU-only and offloads model inference to separate accelerator hardware; AMD’s reference architecture pairs a Helion CPU rack with a Helios GPU rack. The enterprise model merges CPU and GPU blades within one rack, with GPU resources added in groups of 8-GPU MI350P server nodes.

AMD structures the hyperscale rack using three CPU node classes optimized for different agent types. Personal agents run on dense servers with 256-core Venice CPUs operating at 400W. Enterprise agents use balanced servers equipped with 256-core Venice processors rated for 600W. Technical agents rely on high-performance nodes built around 96-core high-frequency Venice silicon. With 256 cores per socket, the maximum 32,768-core rack capacity is achieved using 128 processors inside a single rack.

The enterprise rack also defines three node roles. Agent sandbox nodes run either the dense 256-core, 400W Venice processor or a 96-core Venice part clocked at 5.0GHz. Inference nodes host the Instinct MI350P GPUs, supporting up to 16 GPUs per rack with a combined FP4 compute rating of 74 petaflops. General-purpose nodes deploy Venice on SP7 or SP8 sockets for database and storage workloads, or Venice-X for heavy technical compute tasks. A built-in token router routes agent model requests across frontier and open-source models.

These processor bins align with AMD’s July Venice disclosures, including the 256-core EPYC 9996 and EPYC 9956 (the latter rated at 400W), plus the 96-core high-frequency EPYC 9686F. For GPU host nodes, AMD specifies high-frequency Venice variants paired with RDIMM or MRDIMM memory, alongside a low-power Zen 6 host processor optimized for LPDDR memory. AMD plans to launch this LPDDR-based host chip in the second half of 2027.

“Agentic AI is not a single workload,” stated Dan McNamara, senior vice president and general manager of Compute and Enterprise AI at AMD. McNamara explained AMD is combining EPYC compute performance, open software tooling, and flexible rack designs so customers can build infrastructure tailored to their unique agent workload requirements.

Ecosystem Reference Design Development

AMD notes Helion configurations are currently in development with ecosystem partners, with commercial availability targeted for November 2026. The firm has not yet published full per-rack power budgets, node counts for the hyperscale variant, or pricing details. MiTAC, Inventec, ASUS, ASRock Rack, and GIGABYTE join the broader Helion partner roster. Multiple vendors announced their Helion roadmaps alongside AMD:

• ---Cisco: Jeremy Foster, SVP and GM of Cisco Compute, says agentic AI reshapes infrastructure requirements and stresses end-to-end connectivity, platform management, and system-level security.
• ---Dell Technologies: Dell will integrate Helion within Dell PowerRack, with extended capabilities for PowerRack networking and storage. Arun Narayanan, SVP of Compute and Networking, describes it as “a complete, validated path to deploy agentic compute at scale.”
• ---Hon Hai Technology Group (Foxconn): Hon Hai highlights rack-level system integration and global manufacturing capabilities to serve hyperscale and enterprise deployments.
• ---HPE: Building on its prior Helios deployment agreement, HPE will continue collaborating with AMD on open standards, including air-cooled and direct liquid-cooled networking solutions.
• ---Lenovo: Lenovo will integrate Helion into its Lenovo Hybrid AI Advantage portfolio.
• ---QCT: President Mike Yang states the open blueprint lets QCT flexibly customize compute configurations.
• ---Supermicro: Supermicro will bring Helion to market as an ORv3-compliant rack-scale architecture under its Data Center Building Block Solutions (DCBBS) lineup.
• ---Wiwynn: President and CEO William Lin characterizes Helion as “an open, flexible foundation for agentic AI infrastructure.”

Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!