QSFP vs QSFP28 vs QSFP56 vs QSFP112 vs OSFP Explained

· Updated · 5 min read · Itzhak Finkelstein, T.E.S IT-SOLUTIONS

Three QSFP cable connectors and one wider finned-top OSFP connector lined up on a grey mat

QSFP, QSFP28, QSFP56 and QSFP112 share one 4-lane shape and differ in lane speed: about 14G, 25G, 50G and 100G per lane, which gives 56G, 100G, 200G and 400G links. OSFP is a different cage. Twin-port OSFP carries 8 lanes of 100G-PAM4 (2x400G, sold as 800G) and does not accept QSFP plugs. This guide is for engineers and buyers who need to match InfiniBand and Ethernet cables to switch and adapter ports.

Lanes times lane rate: the only formula you need

Lanes and lane rate per form factorSFP28 1 × 25G = 25GbE; QSFP+ 4 × 14.0625 = 56Gb/s; QSFP28 4 × 25G = 100Gb/s (NRZ); QSFP56 4 × 50G = 200Gb/s (PAM4); QSFP112 4 × 100G = 400Gb/s (PAM4); QSFP-DD 8 × 50G = 400GbE (PAM4); OSFP 4 × 100G = 400Gb/s (PAM4); OSFP twin-port 8 × 100G = 2x400Gb/s (800G).SFP2825G= 25GbEQSFP+4 × 14.0625= 56Gb/sQSFP28NRZ25G25G25G25G= 100Gb/sQSFP56PAM450G50G50G50G= 200Gb/sQSFP112PAM4100G100G100G100G= 400Gb/sQSFP-DDPAM450G50G50G50G50G50G50G50G= 400GbEOSFPPAM4100G100G100G100G= 400Gb/sOSFP twin-port100G100G100G100G100G100G100G100G= 2x400Gb/s (800G)Lanes and lane rate per form factorSFP28 1 × 25G = 25GbE; QSFP+ 4 × 14.0625 = 56Gb/s; QSFP28 4 × 25G = 100Gb/s (NRZ); QSFP56 4 × 50G = 200Gb/s (PAM4); QSFP112 4 × 100G = 400Gb/s (PAM4); QSFP-DD 8 × 50G = 400GbE (PAM4); OSFP 4 × 100G = 400Gb/s (PAM4); OSFP twin-port 8 × 100G = 2x400Gb/s (800G).SFP28= 25GbE25GQSFP+= 56Gb/s4 × 14.0625QSFP28NRZ= 100Gb/s25G25G25G25GQSFP56PAM4= 200Gb/s50G50G50G50GQSFP112PAM4= 400Gb/s100G100G100G100GQSFP-DDPAM4= 400GbE50G50G50G50G50G50G50G50GOSFPPAM4= 400Gb/s100G100G100G100GOSFP twin-port= 2x400Gb/s (800G)100G100G100G100G100G100G100G100G
One block per electrical lane; the port speed is lanes × lane rate.

Every form factor here is a set of electrical lanes. Link speed is lanes multiplied by the rate per lane. The connector name tells you the lane count and the speed class.

  • SFP family: one lane. The NVIDIA QSA28 adapter wires the SFP28 channel to lane 1 of a QSFP28 port and leaves the other 3 lanes unconnected.
  • QSFP family (QSFP+, QSFP28, QSFP56, QSFP112): four lanes.
  • QSFP-DD: eight lanes. The cage also accepts QSFP plugs.
  • OSFP: four lanes (single-port, 400G) or eight lanes (twin-port, 2x400G).

Modulation is the second variable. QSFP28 runs 4x25G-NRZ. QSFP56 moved to 4x50G-PAM4. QSFP112 and OSFP run 100G-PAM4 per lane. The QSFP cage shape stayed the same across these generations.

Comparison table

Form factorLanesPer-lane rateMax speedTypical IB generationExample TES part
SFP28125G25GbENone (Ethernet)MCP2M00-A005
QSFP+ (QSFP)4up to 14.0625 Gb/s56Gb/sFDR (56Gb/s); QDR/FDR10 at 40Gb/sMC2207130-001
QSFP28425G NRZ100Gb/sEDRMCP1600-E00B
QSFP56450G PAM4200Gb/sHDR (HDR100 on split ends)MCP1650-H001E30
QSFP1124100G PAM4400Gb/sNDR (adapter side)MMA1Z00-NS400
QSFP-DD850G PAM4400GbENone (Ethernet only)MCP1660-W001
OSFP single-port4100G PAM4400Gb/sNDR (ConnectX-7 side)MCP7Y00-N001 (400G ends)
OSFP twin-port8100G PAM42x400Gb/s (800G)NDR (switch side)MCP4Y10-N002, MCA4J80-N005

InfiniBand names map to speeds as follows: QDR/FDR10 40Gb/s, FDR 56Gb/s, EDR and HDR100 100Gb/s, HDR 200Gb/s, NDR 400Gb/s.

What plugs into what

Same 100G, different lanesHDR100: 2 lanes of 50G PAM4. 100GbE on 2x50G-PAM4: 2 lanes of 50G PAM4. 100GbE on 4x25G-NRZ: 4 lanes of 25G NRZ. The first two do not link with the third.HDR100PAM450G50G= 100G100GbE (2x50G)PAM450G50G= 100GbE100GbE (4x25G)NRZ25G25G25G25G= 100GbESame 100G, different lanesHDR100: 2 lanes of 50G PAM4. 100GbE on 2x50G-PAM4: 2 lanes of 50G PAM4. 100GbE on 4x25G-NRZ: 4 lanes of 25G NRZ. The first two do not link with the third.HDR100PAM4= 100G50G50G100GbE (2x50G)PAM4= 100GbE50G50G100GbE (4x25G)NRZ= 100GbE25G25G25G25G
HDR100 and 2x50G-PAM4 100GbE are not compatible with 4x25G-NRZ 100GbE: same speed, different lanes.

The QSFP line is backward compatible by design. A faster cage accepts the slower plugs. The link cannot run faster than the slower part.

  • QSFP56 cage: accepts 100G EDR cables and transceivers. 200GbE QSFP56 cages are backward compatible to QSFP28.
  • QSFP112 cage: supports QSFP56 (4x50G-PAM4) and QSFP28 (4x25G-NRZ) devices. On ConnectX-7 and BlueField-3, QSFP112 ports accept 200G QSFP56 and 100G QSFP28 parts.
  • QSFP-DD cage: accepts all QSFP-based cables and transceivers. QSFP-DD cables and transceivers are Ethernet only. ConnectX adapters are not offered with QSFP-DD cages.
  • QSFP28 port to SFP28: use a QSA28 adapter. It takes SFP28, SFP+ or SFP modules and cables. It is qualified for 10GbE SFP+ and 1GbE SFP transceivers.
  • OSFP cage: NVIDIA states that twin-port and single-port OSFP cages are not compatible with the QSFP line. To reach QSFP112 or QSFP56 ports from an OSFP switch, use a splitter cable with an OSFP end and QSFP ends.

Finned-top vs flat-top OSFP

OSFP plugs come in two shells. The difference is where the heat sink sits.

A finned-top and a flat-top OSFP module on a steel bench, a PCIe card with an OSFP cage behind
  • Finned top: the fins are on the plug. NVIDIA Quantum-2 and Spectrum-4 air-cooled switches require finned-top twin-port OSFP shells for extra cooling.
  • Flat top: the plug is smooth and the cooling fins are on the host cage (a riding heat sink, RHS). ConnectX-7 and BlueField-3 use only flat-top shells for single-port OSFP, QSFP112 and QSFP56. ConnectX-7 OSFP cards support the RHS cage only.
  • DGX H100: its internal cages have riding heat sinks, so they take twin-port, flat-top 800G OSFP plugs.
  • -FLT suffix: on the MCP7Y00, NVIDIA marks the flat-top twin-port OSFP version with -FLT, for liquid-cooled systems. Our MCP4Y10-N001-FLT listing follows the same naming.

A typical NDR cable such as the MCP7Y00 splitter shows both: a finned twin-port end for the Quantum-2 or Spectrum-4 switch and two flat-top 400G OSFP ends for ConnectX-7 adapters.

Common mistakes

Twin-port OSFP copper: passive or activeLog-scale ruler from 0.5 m to 5 m. MCP4Y10 passive DAC in 0.5, 1, 1.5 and 2 m. MCA4J80 active copper cable in 3, 4 and 5 m. NVIDIA says to use passive DAC below 3 m.passive DAC below 3 mactive copperMCP4Y10 passive DAC0.5 m to 2 mMCA4J80 active copper3 m to 5 m0.5 m1 m2 m3 m5 mTwin-port OSFP copper: passive or activeLog-scale ruler from 0.5 m to 5 m. MCP4Y10 passive DAC in 0.5, 1, 1.5 and 2 m. MCA4J80 active copper cable in 3, 4 and 5 m. NVIDIA says to use passive DAC below 3 m.passive DAC below 3 mactive copperMCP4Y10 passive DAC0.5 m to 2 mMCA4J80 active copper3 m to 5 m0.5 m2 m3 m5 m
Length range of the two twin-port OSFP copper families (log scale).
  • Treating HDR100 as 100GbE. HDR100 and 100GbE built on 2x50G-PAM4 are not compatible with 100GbE built on 4x25G-NRZ. Same 100G, different lanes.
  • Ordering QSFP-DD for an InfiniBand fabric. QSFP-DD is Ethernet only.
  • Putting a finned-top plug into an adapter. ConnectX-7 and BlueField-3 use flat-top plugs. Finned-top twin-port plugs belong in air-cooled Quantum-2 and Spectrum-4 switches.
  • Expecting OSFP to take a QSFP cable. It does not. Plan a splitter or a different port.
  • Running passive copper too far. The MCP4Y10 twin-port OSFP DAC comes in 0.5, 1, 1.5 and 2 m. The MCA4J80 active copper cable adds an equalizer IC at each end and covers 3, 4 and 5 m. NVIDIA says to use passive DAC below 3 m.

Sources: NVIDIA LinkX guide for 50G-PAM4 and 25G-NRZ cables (PDF); NVIDIA LinkX 100G-PAM4 product line overview; NVIDIA guide to linking DGX, switches, ConnectX-7 and BlueField-3; MCP7Y00 splitter specifications; MCP4Y10 DAC ordering list; MCA4J80 active copper specifications; QSA28 adapter specifications; Mellanox 56Gb/s QSFP+ DAC brief (PDF); MCP1600-E QSFP28 DAC specifications; MCP2M00-A SFP28 DAC specifications; MCP1660-W QSFP-DD DAC specifications; ConnectX-7 specifications.

Parts we stock

Parts in this post

We rebuild and test every cable before it ships. For cables, the final test covers connector and latch inspection, label, part number and length, jacket condition, and an electrical test where a setup exists for that cable type. See the full range on our interconnect page.

FAQ

What is the difference between QSFP28 and QSFP56?

Both have four lanes and fit the same QSFP cage family. QSFP28 runs 4x25G-NRZ for 100Gb/s (EDR or 100GbE). QSFP56 runs 4x50G-PAM4 for 200Gb/s (HDR or 200GbE). A QSFP56 cage accepts QSFP28 EDR cables and transceivers.

Can I plug a QSFP cable into an OSFP port?

No. NVIDIA states that single-port and twin-port OSFP cages are not compatible with QSFP devices. To connect an OSFP switch port to QSFP112 or QSFP56 adapters, use a splitter cable with an OSFP end and QSFP ends.

What does 800G twin-port OSFP mean?

A twin-port OSFP plug holds two 400Gb/s ports in one cage, using 8 lanes of 100G-PAM4. A 32-cage Quantum-2 switch therefore provides 64 ports of 400Gb/s NDR. Cables such as the MCP4Y10 and MCA4J80 use this connector.

Finned-top or flat-top OSFP: which one do I need?

Air-cooled Quantum-2 and Spectrum-4 switches require finned-top twin-port OSFP plugs. ConnectX-7 adapters and DGX H100 cages have riding heat sinks and take flat-top plugs. The -FLT suffix marks a flat-top twin-port version, for example MCP4Y10-N001-FLT.

Does QSFP112 work with older QSFP56 and QSFP28 cables?

Yes. A QSFP112 cage (4x100G-PAM4, 400Gb/s) supports QSFP56 (4x50G-PAM4) and QSFP28 (4x25G-NRZ) devices. The link cannot exceed the rating of the slower part.

Not sure which connector your switch or adapter needs? Send us the part number or the requirement and we will match the cable.

All posts · Shop by product line