TL;DR
- Fifth-generation SXM mezzanine module socket used by H100, H200 and (with refresh) Blackwell GPUs.
- Provides power delivery up to 700 W and NVLink 4.0 (900 GB/s) connectivity per GPU.
- Mechanically and electrically distinct from PCIe — SXM5 GPUs slot into HGX baseboards via dedicated mezzanine connectors.
- Successor SXM6 / next-generation Blackwell baseboards extend power delivery to 1,000+ W.
Overview#
SXM (Server PCI Express Module) is NVIDIA's family of high-power GPU mezzanine modules. Unlike a PCIe card, an SXM GPU plugs into a dedicated socket on an HGX baseboard, with separate power delivery, NVLink routing and cooling. The SXM5 generation, introduced with H100, supports 700 W TDP and NVLink 4.0 at 900 GB/s.
The form factor matters because it is the substrate of every modern HGX system. An HGX-H100 baseboard wires eight SXM5 GPUs through NVSwitch chips into a non-blocking all-to-all fabric; OEMs (Supermicro, Dell, HPE, Quanta and others) drop these baseboards into chassis to produce H100 servers.
Specifications#
| Metric | SXM5 |
|---|---|
| Power delivery | Up to 700 W per GPU |
| NVLink | NVLink 4.0 (900 GB/s per GPU) |
| GPU support | H100, H200, B100, B200 (variants) |
| Baseboard density | 8 GPUs per HGX baseboard |
| Connector | Mezzanine (not PCIe) |
| Cooling | Air or direct-to-chip liquid |
Why a Mezzanine Module?#
PCIe limits power delivery to 600 W in current specifications and cannot route NVLink. SXM addresses both: dedicated power planes deliver hundreds of amps to each GPU, and on-baseboard traces (or copper-direct cables for newer generations) route NVLink between GPUs and NVSwitch chips.
The trade-off is platform specificity. SXM GPUs cannot be moved between chassis without supporting baseboards; OEMs cannot mix and match SKUs across baseboard revisions the way they can with PCIe cards.
When SXM Matters#
- HGX-platform GPUs (H100/H200/B100/B200 SXM variants).
- Multi-GPU training and high-throughput inference where NVLink bandwidth dominates.
- Liquid-cooled deployments where direct-to-chip cooling sockets onto SXM modules.
- Pick PCIe variants for retrofit servers or single-card workloads.
Pitfalls#
- Cross-vendor compatibility is limited — SXM baseboards are platform-specific.
- Cooling provisioning is non-trivial; SXM5 at 700 W needs air-cooled rear-door heat exchangers at minimum.
- Maintenance is more involved than PCIe — swapping an SXM module requires baseboard access.
- Next-generation Blackwell baseboards push power envelopes above 1 kW per GPU; SXM5 chassis cannot accommodate them without redesign.
Software Notes#
SXM is transparent to the CUDA software stack. Drivers, NCCL topology files and tooling treat SXM and PCIe GPUs identically except for capability flags (NVLink topology, power-cap profiles).
References
- NVIDIA HGX Platform Overview · NVIDIA