A rack-scale computing platform aimed at artificial intelligence agent workloads is taking shape at AMD, with hardware configurations spanning personal, enterprise, and technical use cases. Carrying the Helion Agentic AI Rack designation, the system accommodates a ceiling of 32,768 Zen 6 x86 CPU cores alongside 16 Instinct MI350P accelerators, with an FP4 throughput capacity reaching 74 petaflops.
Networking throughput on the platform reaches 102 Tbps, while a modular hardware architecture accommodates varying computational demands across different agent workloads. Three distinct workload types — personal assistants, enterprise applications, and technical computing — sit at the heart of AMD's framework for agent tasks, with each category carrying its own set of processing demands.
Compact, power-efficient server hardware sits at the heart of what personal agents demand, whereas enterprise deployments call for a measured balance of agent throughput and overall performance, and technically intensive workloads place the greatest weight on raw processing speed. Operators gain the flexibility to configure CPU core density, processing throughput, and power draw to match the demands of their specific workloads through AMD's configuration options.
Two distinct hardware philosophies separate the product tiers: the hyperscale build draws entirely on central processing units, whereas the enterprise rack bundles those processors alongside graphics accelerators within a single integrated enclosure. Handling personal-agent workloads falls to 256-core Venice processors drawing 400W, a configuration chosen to balance computational throughput against energy consumption.
Two distinct Venice processor variants serve different workload demands: the 256-core, 600W model handles enterprise tasks, while a 96-core, higher-frequency version is deployed for technical computing applications. Under that enterprise setup, agent execution, inference workloads, and general-purpose computing each run on their own dedicated server nodes rather than sharing infrastructure.
Three distinct hardware layers divide the workload: sandbox nodes built around CPUs take on agent execution, Instinct MI350P accelerators are dedicated to inference tasks, and a tier of general-purpose servers absorbs the demands of databases, storage, and technical computing. A built-in token router ships with the platform, handling the distribution of agent model requests between frontier and open-source options.
Networking within the rack runs on a programmable basis, drawing on an open-source software ecosystem, while the hardware itself — 44U units built to ORv3 specifications — conforms to the standards set out by the Open Compute Project.
Racks configured to hold up to 32,768 Zen 6 CPU cores alongside dense GPU arrays fall well outside what conventional data centre infrastructure — whether cloud-oriented or built around older enterprise workloads — was designed to handle. Data centres retrofitted around pre-existing server layouts often fall short of the power distribution, thermal management, and connectivity demands those hardware arrangements place on the surrounding infrastructure.
Three distinct computing categories — personal agents, enterprise applications, and technical workloads — each place fundamentally different demands on processing infrastructure, compounding the difficulty of managing a varied task portfolio. Supporting two processor families that differ in core count and clock behaviour — 256-core dense units on one side, 96-core high-frequency variants on the other — the rack layout AMD has proposed draws non-trivial infrastructure requirements into focus, given that each server type brings its own power, cooling, and management footprint to the floor.
Juggling rack space, power draw, and cooling capacity across varied server hardware — rather than rolling out uniform equipment throughout — leaves data centre operators with a layered infrastructure puzzle to solve. Where hardware cannot handle the pairing of dense-core processors with high-frequency clock speeds, those running the network may face an obligation to retrofit or replace the underlying equipment.
On the networking side, the rack handles up to 102 Tbps of bandwidth while housing a programmable token router — a combination that brings its own set of design factors into play. Networking hardware and software at existing sites may require upgrades or replacement to handle that level of throughput and direct model requests across multiple AI platforms.
Planned for release in November 2026, the Helion Agentic AI Rack arrives without AMD having made public the full costs tied to facility upgrades or the cooling demands the hardware will place on data centre infrastructure.
