Two acronyms decide how an AI or storage server is physically built, and both confuse buyers who are otherwise comfortable reading a spec sheet. SXM versus PCIe is the question of how a GPU plugs into the system and how GPUs talk to each other. EDSFF, with its E1.S and E3.S variants, is the new ruler-shaped form factor quietly replacing the U.2 NVMe drive. Get either wrong and you buy a chassis that cannot deliver the bandwidth, density or cooling the rest of your investment assumes. This is the what-is-it layer that sits underneath the GPU and SSD buying decisions, so the comparison tables make sense before you spend.
SXM vs PCIe: two ways to fit a GPU
A PCIe GPU is a card. It slots into a standard PCIe x16 slot, the same socket any expansion card uses, and it is the flexible, universal option: it drops into a mainstream rack server, draws power within the slot and supplementary connectors, and is straightforward to add, remove or mix. NVIDIA H200 PCIe and L40S are examples of this format, and it is the right starting point for most teams, which is why it anchors our GPU accelerators range.
SXM is not a card at all. It is a mezzanine module that bolts directly onto a dedicated baseboard inside a purpose-built GPU server. Because it is not constrained by the PCIe slot, an SXM module can draw far more power and, crucially, connects to its neighbours over NVLink and NVSwitch, a high-bandwidth GPU-to-GPU fabric that a PCIe card cannot fully use. That is the whole point of SXM: it exists to let eight GPUs behave more like one very large accelerator.
Why the difference matters: bandwidth, power and cooling
For a single GPU, or a few that work independently, PCIe is more than adequate and far more flexible. The gap opens when GPUs must share data constantly, as they do when training or serving a large model split across many devices. There, the NVLink fabric between SXM modules moves data between GPUs at bandwidth well beyond what PCIe lanes provide, and that interconnect, not the raw compute of any one GPU, is what keeps a large job from stalling.
Power and cooling follow from this. SXM modules run at substantially higher per-GPU power than slot-limited PCIe cards, which is why SXM systems are dense, heavily engineered and increasingly liquid-cooled, while PCIe GPUs remain serviceable in conventional air-cooled racks. The trade is real on both sides: SXM delivers maximum multi-GPU performance and density at the cost of flexibility and cooling complexity; PCIe trades peak scale-up bandwidth for the ability to live in an ordinary server. We weigh that against the cooling envelope in AI server cooling, air vs liquid.
- •PCIe GPU: standard slot, flexible, air-coolable, ideal for single or independent GPUs
- •SXM module: baseboard-mounted, higher power, NVLink/NVSwitch fabric for tight multi-GPU work
- •NVLink bandwidth between SXM GPUs far exceeds PCIe for shared-memory training
- •SXM density usually implies liquid or heavily-engineered cooling
EDSFF: the ruler that replaces U.2
Storage has its own form-factor shift. For years, enterprise NVMe meant the U.2 drive, a 2.5-inch device that inherited the footprint of the old SAS and SATA disks it replaced. That footprint was designed for spinning platters, not flash, and it limits how many drives fit across the front of a server and how well they cool.
EDSFF, the Enterprise and Data Center Standard Form Factor, redesigns the drive around flash from scratch. Instead of a square, it is a thin ruler that stands on edge, so many more drives line up across a chassis front, airflow runs cleanly along each device, and capacity and bandwidth per rack unit rise sharply. EDSFF is the direction modern all-NVMe servers are taking, and it underpins how we plan dense storage in SSD and NVMe.
E1.S vs E3.S: which ruler
EDSFF comes in two main flavours, and the names describe their shape. E1.S is the shorter, thinner ruler, optimised for maximum drive count in dense, often 1U servers; it is where hyperscale-style compute nodes pack large numbers of NVMe devices into a slim chassis. E3.S is the larger ruler, roughly a modern successor to U.2 in role, offering higher per-drive capacity and power headroom, and it is becoming the mainstream choice for general-purpose all-flash storage and database hosts.
The practical guidance is simple. E1.S when the goal is the highest number of drives in the least space, typically in scale-out nodes. E3.S when you want a balance of high capacity, performance and serviceability for conventional storage and database servers. Either way, the chassis is built for one standard, so the form factor is a decision you make up front, alongside the drive endurance and capacity choices in SSD and NVMe.
Putting the two together
These choices are not independent. An SXM GPU server is a dense, high-power, often liquid-cooled platform, and it pairs naturally with EDSFF NVMe scratch storage feeding the GPUs at high bandwidth; the whole chassis is engineered as one thermal and electrical system. A PCIe-GPU server in a standard rack is more likely to mix conventional NVMe with a smaller number of accelerators. The form factors travel together because they reflect the same underlying decision: how dense, how fast and how specialised the box needs to be.
If you are comparing specific GPUs, the buyer-level comparison still matters, for example our look at the H200 PCIe. This article is the layer beneath it: once you know whether you need SXM or PCIe, and E1.S or E3.S, the spec sheets stop being a wall of acronyms. We design the full platform, GPU and storage form factors included, in our GPU accelerators practice.