A transceiver form factor defines the mechanical and electrical interface between the pluggable module and the switch port. Not all ports accept all transceivers. Getting this wrong before you order hardware means transceivers that physically cannot be installed.
Transceiver form factors — select to compare
Form factor determines which cage a module fits in and which devices are compatible
No backward compatibility — purpose-built cage only accepts OSFP 400G modules.
The form factor hierarchy
QSFP28 supports 100G using four lanes at 25G each (NRZ modulation). This was the dominant form factor through the 100G era. A QSFP28 port is electrically identical to a QSFP+ port (40G) -- full backward compatibility.
QSFP56 supports 200G using four lanes at 50G each (PAM-4). Some 200G switch deployments use this. Backward compatible with QSFP28 and QSFP+.
QSFP-DD (Double Density) supports 400G or 800G using eight lanes. The extra four lanes (vs QSFP56) come from a second row of electrical contacts in the connector -- the "double density" in the name. QSFP-DD is backward compatible with QSFP28, QSFP56, and QSFP+. This backward compatibility is its primary advantage -- a switch bought today with QSFP-DD ports can accept 100G modules from existing inventory while being ready for 400G and 800G upgrades.
OSFP (Octal Small Form-factor Pluggable) was designed specifically for the 800G era and higher. It is physically larger than QSFP-DD and has a higher power budget (roughly 20W vs 12W for QSFP-DD). OSFP is not backward compatible with QSFP form factors. OSFP-XD (OSFP Extended Density) is the 2026-era form factor for 1.6T links.
The DGX H100 port identity
Every DGX H100 node has eight OSFP ports -- one per ConnectX-7 NIC. The switch connecting these nodes (the SN5600 leaf switch in NVIDIA's reference design) also has OSFP ports. This is not a coincidence -- NVIDIA designed the H100 DGX system and the SN5600 to use the same form factor, enabling a single cable type to connect server to switch across all eight rails.
This also means that when NVIDIA says "the same OSFP ports are used for both InfiniBand and Ethernet operation" -- they mean the same physical port, same transceiver, same cable. The mode is set in firmware on the ConnectX-7. This is the physical reality underlying Chapter 0's "same cable for IB and RoCEv2" statement.
The heat sink is a compatibility axis, not a finish
Two OSFP modules can be the same form factor, the same speed, and the same optics, and still not be interchangeable. What separates them is the top of the module.
| Top | What it is | Where it goes |
|---|---|---|
| Finned-top (IHS) | Integrated heat sink, fins on the module | Switches only -- Quantum-2, SN5600 |
| Flat-top | No heat sink on the module; the cage provides it | Host adapters only -- ConnectX-7 OSFP |
| Flat-top RHS | Riding heat sink; the cage's sink rides on the module | DGX H100 Cedar links, and liquid-cooled designs |
A finned module physically will not seat in a host cage that expects a flat top, and a flat-top module in a switch cage has no thermal path. This is one of the more common "the replacement optic does not fit" tickets, and it is invisible on a datasheet line that only lists speed and reach.
The DGX H100's Cedar links are the case worth remembering: they take flat-top RHS parts specifically, and neither the switch part nor the ordinary HCA part is a substitute.
Reading the cable before you touch it
The jacket colour is a real convention, not decoration:
- Black -- direct attach copper
- Blue -- multi-mode, short reach
- Orange -- single-mode, long reach
From an aisle, that is enough to spot a single-mode module in a rack that should be all copper, which is usually the tell for a cable that was replaced in a hurry with whatever was on the shelf.
Transceiver power, by part
The table above gives ranges by form factor. When you are doing the actual arithmetic for a rack, the per-part numbers are what you want:
| Class | Reach | Module power |
|---|---|---|
| Passive DAC | 1-3 m | under 0.1 W |
| ACC (linear active copper) | 3-5 m | 1.5 W |
| AOC | ~30 m | under 12 W |
| Multi-mode, twin-port 800G | 50 m | 15 W |
| Multi-mode, single-port 400G | 50 m | 8 W |
| Single-mode, twin-port 800G | 100 m - 2 km | 17 W |
| Single-mode, single-port 400G | 100 m | 9 W |
The shape that matters: a twin-port 800G module costs roughly twice the power of the single-port 400G part beside it, and copper costs two orders of magnitude less than either. A row of racks re-planned from DAC to optics because someone moved the leaf switch does not just cost money; it changes the cooling load.
Power budgets and thermal consequences
Form factor choice directly determines power consumption per port, which determines switch total power, which determines rack cooling requirements:
| Form factor | Speed | Module power |
|---|---|---|
| QSFP-DD | 400G | 3-4W per port |
| QSFP-DD | 800G | 8-10W per port |
| OSFP | 400G | 5-8W per port |
| OSFP | 800G | 12-15W per port |
| OSFP-XD | 1.6T | 20-25W projected |
A 64-port 800G switch with OSFP modules: 64 x 15W = 960W in optics alone. Add ~500W for the switch ASIC on a 51.2 Tbps chip, and the total switch power approaches 1,500W -- from a device roughly the width of a pizza box.
Eight such switches in a DGX BasePOD compute fabric: 12,000W from switching alone. This is before accounting for the 32 DGX nodes, storage switches, management networking, or out-of-band infrastructure. Optics are not a minor thermal contribution. They are a primary input to rack cooling design.