AI hardware comes wrapped in form-factor jargon that hides real engineering trade-offs. Should a GPU be SXM or PCIe? What is EDSFF and why are flash drives suddenly shaped like rulers? These are not cosmetic distinctions; they decide bandwidth, power, cooling and how many devices fit in a chassis. This explainer translates SXM, PCIe and EDSFF into plain terms so you can read an AI server specification and understand what each choice actually buys.
SXM vs PCIe: two ways to mount a GPU
PCIe GPUs are cards that slot into the same expansion bus as any other peripheral. They are flexible, broadly compatible and fit mainstream servers, which makes them the natural choice for inference, smaller training jobs and mixed-use boxes. The trade-off is that GPU-to-GPU communication runs over the PCIe bus or a limited bridge, which becomes a ceiling when many GPUs must work together tightly.
SXM is a socketed form factor where GPUs mount onto a baseboard and connect through a dedicated high-bandwidth interconnect rather than the PCIe bus. That gives far higher GPU-to-GPU bandwidth, which is what large multi-GPU training needs, along with higher power and tighter cooling demands. SXM lives in purpose-built eight-GPU platforms; PCIe spreads across the whole server range. Our GPU accelerator guidance covers where each sits.
Bandwidth, power and cooling differ
The practical differences follow from how the GPUs talk to each other and how much power they draw. SXM parts run at higher power envelopes and rely on the baseboard's dedicated interconnect for fast collective operations across all GPUs, which is decisive for training large models where the GPUs exchange data constantly. That power and bandwidth come with serious cooling requirements, often pushing toward liquid in dense configurations.
PCIe parts run at lower power, cool more easily and drop into standard servers, but their inter-GPU bandwidth is more limited. For inference and graphics, where each GPU largely works alone, that limit rarely bites and the flexibility wins. For tightly-coupled training across many GPUs, the SXM interconnect is the reason those platforms exist. Match the form factor to whether your GPUs work together or independently.
What EDSFF means and why it matters
EDSFF, the Enterprise and Data Centre Standard Form Factor, is the modern shape for server flash, replacing the older 2.5-inch U.2 design with formats often described as rulers. The two you will meet most are E1.S, a compact format for dense flash, and E3.S, which is becoming the default for PCIe Gen5 NVMe in current servers. The shapes are not arbitrary: they are engineered for better airflow, higher density and the thermals that fast Gen5 drives demand.
For an AI server this matters because feeding GPUs needs a lot of fast local flash, and EDSFF packs more of it into a chassis while keeping it cool and serviceable. When a current platform specifies E3.S bays, it is signalling a Gen5, high-density flash design rather than a legacy layout. Our SSD and NVMe guidance covers the form factors in detail.
- •PCIe GPUs: flexible, lower power, mainstream servers - best for inference and mixed use
- •SXM GPUs: socketed, high inter-GPU bandwidth, high power - best for large training
- •EDSFF E1.S and E3.S: the modern ruler form factors for dense, cool Gen5 flash
- •E3.S is becoming the default for PCIe Gen5 NVMe in current servers
Reading an AI spec with confidence
Put together, these three terms tell you a lot at a glance. An eight-GPU SXM platform with E3.S NVMe and liquid-ready cooling is a training machine; a server with a few PCIe GPUs and standard NVMe is built for inference or mixed work. Knowing which is which stops you buying a training-grade chassis to serve a model, or trying to train a large model on inference-shaped hardware. For the GPU model choices that sit on top of the form factor, see our H200 PCIe page, and bring the workload to our GPU accelerator guidance to size it correctly.