GPU Form Factors Explained: PCIe, SXM, OAM, and the Variants In Between

A "GPU" on a spec sheet is usually a chip plus a board plus a way of plugging into a system. That last part — the form factor — quietly shapes how a card cools itself, how much power it can pull, how it talks to other GPUs, and where you can actually put it.

If you're shopping for a gaming or content-creation GPU, the form factor question is mostly about slot width and power: every consumer card is a PCIe add-in card. The interesting form factors — SXM, OAM, NVLink-bridged PCIe — only show up on datacenter and AI accelerators, where multi-GPU bandwidth and power density really start to matter.

Below is a quick tour of the form factors used by every GPU currently in the GPU Poet database, with examples you can click through.

PCIe add-in card

The familiar one. A board with a PCIe x16 edge connector that drops into a desktop or server slot. Powered by the slot (up to 75W) plus 6/8/12-pin auxiliary connectors. Cooled either by fans built into the card itself ("active" cooling, what you'd find on a typical desktop GPU) or, in servers, by the chassis fans blowing air through a fanless heatsink on the card ("passive" cooling — only practical with the high-static-pressure airflow a server case provides).

PCIe add-in cards come in different physical sizes that matter when you're picking a card:

• Low-profile, single-slot, slot-powered — short bracket, fits 1U servers and small-form-factor desktops. ≤75W so no auxiliary power connector, often passively cooled. Examples: NVIDIA T4, NVIDIA L4, RTX A1000.
• Compact dual-slot, slot-powered — half-height or full-height with a low-profile bracket option. Still ≤75W, but two slots wide for a bigger cooler. Examples: RTX PRO 2000 Blackwell, Intel Arc Pro B50.
• Full-height, dual-slot — the default for mainstream gaming and most workstation cards. Examples: RTX 5070, Radeon RX 9070 XT.
• Triple-slot and larger — needed for 300W+ cards. Examples: RTX 4090, RTX 5090.
• Full-height full-length, fanless (FHFL passive) — the server form for datacenter PCIe cards. The card has no fan of its own — just bare heatsink fins — so it relies on the server chassis fans pushing high-pressure air across it. Won't cool itself in a regular desktop case. Examples: NVIDIA A100 PCIe, H100 PCIe, L40S.

SXM (Server PCI Express Module)

NVIDIA's mezzanine module format for datacenter GPUs. Instead of a slot, the GPU bolts onto a socket on the server's HGX¹ or DGX² baseboard. The socket carries PCIe to the host CPU plus high-bandwidth NVLink lanes to the seven sibling GPUs on the same board. SXM modules can run at higher TDPs (700W–1000W+) than PCIe cards because the baseboard handles power and cooling as a unit.

The socket has versioned with each generation: SXM2 (Pascal/Volta), SXM3 (Volta refresh), SXM4 (Ampere), SXM5 (Hopper), SXM6 (Blackwell). Modules are not slot-compatible across versions.

Examples in the database: Tesla P100 (SXM2), Tesla V100 16GB and 32GB (SXM2), H200 SXM (SXM5), B100 and B200 (SXM6).

Servethehome has a good walk-through of installing SXM2 modules — torque-spec details and all — at How to Install NVIDIA Tesla SXM2 GPUs in DeepLearning12. The video below shows the physical assembly clearly.

OAM (OCP Accelerator Module)

The Open Compute Project's vendor-neutral answer to SXM. An OAM is a 102mm × 165mm mezzanine module that bolts onto a Universal Base Board (UBB³) holding up to eight accelerators. The UBB carries PCIe to the host and a fabric (Infinity Fabric on AMD, scale-up Ethernet on some designs) between the accelerators. Because OAM is an open standard, the same socket can host parts from different vendors.

AMD's Instinct datacenter line uses OAM exclusively. Examples: MI300X, MI325X, MI350X, MI355X. The OCP specification is published at opencompute.org.

NVL (PCIe with NVLink bridge)

A variant on the FHFL passive PCIe card: each card has NVLink bridge connectors at the top edge so two or four cards can be linked into a single pooled-memory GPU group for LLM inference, without giving up the standard PCIe form factor. Example: NVIDIA H200 NVL — 141GB HBM3e per card, 282GB pooled in a 2-way bridge or 564GB in a 4-way bridge.

Workstation vs. server variants of the same chip

Some chips ship in multiple board variants that share a socket family (PCIe) but differ in cooling, TDP, and bracket. The RTX PRO 6000 Blackwell, for instance, ships as a Workstation edition (built-in fan, 600W), a Server edition (no fan on the card — cooled by chassis airflow, 600W), and a Max-Q variant (built-in fan, 300W for multi-GPU workstations). Same chip and feature set, different physical card. The RTX PRO 4500 has workstation and server editions for the same reason.

MXM and mobile

Not currently in the database, but worth knowing: MXM is the mezzanine standard for laptop and embedded GPUs. If you see a "mobile" or laptop variant of a GeForce or Radeon card, it's typically an MXM module or a directly-soldered BGA part, with TDPs and clocks much lower than the desktop card of the same name.

Practical buying notes

• A "PCIe" version of a datacenter GPU is usually slower than the SXM/OAM version of the same chip, because the socketed variant gets a higher power budget and far more inter-GPU bandwidth. Some PCIe cards do have NVLink bridge connectors (the H200 NVL above, plus A100 PCIe and H100 PCIe), but a bridge only links 2–4 adjacent cards. SXM modules on an HGX¹ baseboard get all-to-all NVLink across all 8 GPUs through an NVSwitch fabric, at much higher aggregate bandwidth.
• SXM and OAM modules can't be dropped into a regular PC — they need the matching baseboard (HGX¹ or DGX² for SXM, UBB³ for OAM).
• Server PCIe cards (the fanless "passive" datacenter variants) won't cool themselves on a desk; they need a server chassis pushing serious airflow.
• Slot width and TDP determine whether a card will physically fit and whether your PSU can feed it.

More background reading: NVIDIA's HGX platform overview, the OCP OAM specification, and the Servethehome SXM2 install guide linked above.