OAM (Open Accelerator Module) Form Factor

Overview#

OAM — Open Accelerator Module — is the Open Compute Project's standardised mezzanine module specification for high-power accelerators. It exists to provide a vendor-neutral alternative to SXM: any accelerator vendor can ship parts in the OAM form factor, and any baseboard vendor can build chassis that accept them.

AMD's Instinct family (MI250, MI300X, MI325X, MI355X), Intel Gaudi 2 and 3, and a long list of specialised accelerators ship OAM modules. The associated Universal Baseboard (UBB) spec defines an 8-module reference platform that OEMs can build to.

Specifications#

Why an Open Standard?#

SXM is proprietary; only NVIDIA accelerators can occupy SXM sockets. The hyperscaler customers who drove OCP's formation wanted a way to source accelerators from multiple vendors without redesigning chassis for each.

OAM is the result. The mechanical, electrical and thermal envelopes are standardised; the inter-module fabric (e.g. AMD Infinity Fabric, Intel RoCE) is vendor-specific. Customers can build OAM-platform chassis once and use them across multiple accelerator generations and vendors.

When OAM Matters#

• Multi-vendor accelerator strategies (AMD + Intel + others in same form factor).
• Hyperscaler procurement where supply diversification and chassis reuse matter.
• Sovereign AI builds that prefer open standards over proprietary stacks.
• For pure NVIDIA buys — SXM is the de facto path.

Pitfalls#

• Inter-module fabrics differ by vendor — code optimised for Infinity Fabric does not run on Gaudi's RoCE without changes.
• Cooling profiles vary; not all UBB baseboards support all module TDPs.
• Software stack maturity varies dramatically across OAM vendors — ROCm has improved markedly, others lag.

Software Notes#

OAM is transparent to operating systems and drivers. The accelerator stack (ROCm, SynapseAI, vendor-specific) is what matters; OAM is purely a hardware integration standard.