OAM vs SXM

OAM, the OCP Accelerator Module, is an open mezzanine form factor for AI accelerators, defined by the Open Compute Project. SXM is NVIDIA's own module format for its data center GPUs. In eight-accelerator servers, OAM modules sit on the OCP Universal Baseboard (UBB), and SXM modules sit on NVIDIA's HGX baseboard. OCP states that the UBB is form factor compatible with the HGX baseboard, so a chassis can be designed around either.

The UBB 2.0 base specification supports OAMs up to 1,000 W each and 12 kW per baseboard, though individual designs may differ. AMD rates its MI355X at 1,400 W per module, and each carries 288 GB of HBM3E high-bandwidth memory. NVIDIA configures each B200 SXM GPU at up to 1 kW. The fabrics differ too. Each MI355X has seven direct Infinity Fabric links, one to each peer, at 153.6 GB/s bidirectional per link. On HGX B200, NVSwitch chips give each GPU 1,800 GB/s of GPU-to-GPU bandwidth.

Choose OAM for AMD Instinct MI350 Series parts and other OAM modules, and SXM for NVIDIA GPUs that need NVSwitch bandwidth. Size rack power to the module's own rating, not the spec's 1,000 W figure, since eight 1,400 W MI355X modules draw 11.2 kW before CPUs and fans.

Sources

Last verified October 7, 2026.