Skip to main content
AMD Helion Racks: 73,728 Threads, 3 Variants, 0 Prices - shattered.io

AMD Helion Racks: 73,728 Threads, 3 Variants, 0 Prices - shattered.io

AMD Processors
16 lis 2026, 07:47Hardware News
Oryginał
3 min read

AMD announced general availability of its Helion rack family on November 10, 2026, offering three variants for agent workloads with thread counts reaching up to 73,728 per rack.

AMD turned the Helion concept into buyable hardware on November 10, 2026, with three named deployment options for agentic AI workloads. The company declared general availability of its 6th Gen EPYC server processors on the SP7 platform, which power the new Helion racks. Server makers began shipping systems based on the EPYC 9006 Venice chips on the same day.

Helion is an open blueprint for agent infrastructure, built on an Open Compute Project specification and designed to run alongside AMD’s separate Helios rack system. Helios targets GPU-heavy training and inference, while Helion focuses on CPU-led agent execution. The two families are meant to operate side by side in a single data center.

Three variants for different cooling needs

AMD offers three Helion variants that differ in cooling and GPU content. Helion Liquid is the high-density option, designed for maximum agent capacity per rack. It relies on liquid cooling to support the highest thread counts AMD advertises. Helion Air targets the installed base of air-cooled data centers, while Helion Air Inference combines CPU-based agent execution with local GPU inference.

The naming reflects the physical constraints of data center infrastructure. Liquid-cooled halls can support the densest configurations, while air-cooled racks hit a power ceiling long before reaching the top thread counts. Helion Air Inference places GPUs a few inches from the CPUs to reduce network latency for model calls.

Thread counts and power demands

AMD claims some high-density Helion configurations reach up to 73,728 threads per rack. This figure comes from AMD’s launch materials and is corroborated by TrendForce. The top configuration likely uses the EPYC 9996 processor, which has 256 cores and 512 threads with SMT enabled. A 73,728-thread rack would require 144 such processors.

The power requirements for such a dense build are substantial. AMD’s footnotes model a 35 kW air-cooled rack with a maximum of 20 two-socket servers, which yields 20,480 threads. The 73,728-thread configuration carries more than three times that load, explaining why Helion Liquid exists. A 256-core SP7 part can consume 400 W to 600 W, pushing the CPU power alone to roughly 57.6 kW to 86.4 kW.

AMD’s own materials cite a separate design point of up to 32,768 CPU cores, or 65,536 threads, for a CPU-only hyperscale configuration. This figure likely describes a different build than the 73,728-thread ceiling. Until AMD reconciles the two numbers, the 73,728 figure should be treated as a maximum for the densest liquid-cooled rack.

Open design and market positioning

Helion follows an Open Compute Project specification, allowing multiple vendors to build compatible racks. This openness contrasts with sealed, single-vendor appliances. Helios, the GPU-focused rack, is based on Meta’s 2025 OCP Open Rack for AI design. The choice of open standards lets ecosystem partners compete on configuration and integration.

The market impact of Helion remains to be seen. AMD’s materials list a two-socket EPYC 9996 setup at $14,904 against a two-socket Xeon 6990E+ at $14,995, but these are processor prices, not rack prices. No official list price exists for Helion Liquid, Helion Air, or Helion Air Inference. Procurement teams will need to obtain quotes from OEMs and ODMs based on their specific configurations.

Cloud availability is expected to follow the hardware. AMD stated that Microsoft Azure and Oracle Cloud Infrastructure instances on the new chips are coming soon, but it did not provide specific dates. That ambiguity matters for anyone planning capacity, as a coming-soon label can mean weeks or quarters.

Z.ai
Article written by AI
Z.ai GLM-4.7-FlashFree
Free AI articles150 of 150
No limit after signing in with Telegram