The Helios AI is AMD's massive, double-wide GPU-focused compute blueprint that has eighteen blades with four GPUs and one CPU each. It's an incredibly powerful system for training frontier AI models like those used for ChatGPT, but most customers aren't doing that. The hot new thing now is agentic AI, and AMD's new Helion platform is directly targeted at those workloads.

Two views of AMD Helion next to the larger Helios rack. - Image: AMD

Helion is smaller than Helios, being built in the standard OpenRack v3 form factor instead of the Open Rack Wide (ORW) form factor created by Meta. Both are different from classic 19-inch racks; OpenRack is 21" wide inside and roughly 24" outside; ORW is 48" wide outside, so it is essentially double-wide. OpenRack v3/v4 happens to be the same standard NVIDIA uses for its NVL72 rack systems, as the MGX architecture those are built on is also part of OCP.

AMD Helion General Specifications - Image: AMD

Helion can fit in the smaller racks because it is targeted at agentic workloads, which are much less GPU-heavy than familiar generative AI tasks like image generation. While those kinds of tasks are compute-hungry, agentic workloads require considerable CPU horsepower as there are a lot of non-AI components to the workload. Even NVIDIA acknowledges that the ideal mix of CPUs to GPUs for agentic AI is something like 1:1.

AMD Helion Enterprise Configuration Examples - Image: AMD

As a result, Helion is available in both Enterprise and Hyperscale configurations. The Enterprise configuration mixes EPYC Venice CPUs with AMD Instinct MI350P accelerators as well as up to 102 Tbps networking bandwidth, which we're taking as a scale-out number.

AMD Helion Hyperscale Configuration Examples - Image: AMD

Meanwhile, the Hyperscale configuration is meant to be paired with a Helios GPU rack, and it comprises up to 128 EPYC CPUs, supplying up to 32,768 Zen 6c CPU cores, although AMD also notes that it's possible to equip the machine with 96 "Venice HF" chips that are designed to hit high frequency for agents that are performing highly technical tasks and need the lowest latency possible.

Ultimately these are just examples, though; you're actually free to configure Helion however you want, or at least, however you can get your B2B salesperson to sell it to you. AMD specifically says you can "choose any combination of CPU and GPU blades". Everything is open-source and open standards, so in theory the configurations should be fairly malleable even after deployment.

AMD is shipping much of its work on Helion back upstream to OCP. - Image: AMD

AMD says it plans to contribute the work it has put into the Helion rack specification back to OCP, much as NVIDIA and Meta have also done with their AI server developments. That means the physical form factor and mounting hardware, not the chips or interconnects, obviously. The firm says it plans for Helion to be available from its ecosystem partners next month.