What is SXM5?: rawcompute.in Glossary

SXM5 is the fifth generation of NVIDIA’s Server PCI Express Module form factor. Unlike a standard PCIe add-in card, an SXM5 GPU module mounts directly onto a baseboard (such as the HGX A100 or HGX H100) via a high-density connector. This design eliminates the PCIe edge connector bottleneck and allows for higher TDP. The H100 SXM5 operates at up to 700 W compared to 350 W for the PCIe variant. The SXM5 socket also carries NVLink signals, enabling the 900 GB/s GPU-to-GPU interconnect that PCIe cards cannot physically support.

An HGX baseboard typically hosts eight SXM5 modules connected through NVSwitch, creating an all-to-all GPU mesh. The baseboard then connects to the host CPU via PCIe Gen5 x16 lanes per GPU. This architecture is the standard building block for large-scale AI training clusters worldwide, including those deployed by hyperscalers and sovereign AI projects.

Why it matters when buying hardware

SXM5 GPUs cost more than their PCIe counterparts and require specific baseboard and chassis combinations (e.g., Supermicro SYS-821GE or Dell PowerEdge XE9680). However, they deliver substantially more performance per GPU for training workloads due to higher TDP and NVLink connectivity. If your workload is multi-GPU training, always specify SXM5. If you are deploying single-GPU inference instances, PCIe variants may offer better cost efficiency. Rawcompute.in can help you pick the right form factor for your workload profile.