OSFP vs QSFP-DD vs QSFP112: Which Should You Buy?

Comparing OSFP, QSFP-DD and QSFP112 as if they were three interchangeable 400G transceivers is the fastest way to buy the wrong part. They are host interfaces first and optical modules second. The port on your switch, network adapter or DPU decides which of the three is even a candidate; reach, fiber and connector are separate decisions made afterwards.

• QSFP-DD - eight host lanes, double-density QSFP mechanics. Pick it when you are migrating inside an existing QSFP estate and need one platform to carry 100G, 200G, 400G and 800G ports.
• OSFP - eight host lanes, a larger body with more thermal room. Pick it when the platform is already OSFP, or when 800G/1.6T and higher-power optics are on the roadmap.
• QSFP112 - four host lanes at roughly 112 Gb/s each. Pick it when the adapter or DPU explicitly provides a QSFP112 cage. It is an endpoint form factor, not a switch form factor.

Mechanical fit tells you almost nothing. A module that slides into a cage and latches can still stay dark because the host runs the wrong lane mode, the firmware does not recognise the module ID, or the port cannot supply the module's power class.

The Host Port Decides Everything

Work left to right through this table. Each column depends on the one before it, and none of them can be skipped.

Notice what the table does not contain: a column for "400G". Aggregate speed is the least useful thing you know about a link. A 400G OSFP transceiver, a 400G QSFP-DD transceiver and a 400G QSFP112 transceiver carry identical traffic and are not substitutes for each other.

Can a QSFP112 Module Work in a QSFP-DD Port?

Sometimes. Never assume it.

The mechanical part of the answer is settled. The QSFP112 module, cage and connector are defined in SNIA's SFF-TA-1027 and are mechanically backwards compatible with the wider QSFP family, and the QSFP-DD MSA folded its QSFP112 electrical and management timing work into the QSFP-DD hardware specification before those clauses were transferred to SNIA SFF. So yes, the module will physically go in and latch.

What happens next is a platform question, and it splits into three:

1. Does the host support a four-lane 100G/112G-class port mode at all? An earlier QSFP-DD switch built around eight 50G lanes reaches 400G by a completely different electrical route. Its SerDes may have no 4×100G mode to configure.
2. Will the firmware admit the module? The port has to read the module ID, apply a management page, negotiate FEC and allow the module's power class.
3. Is the optical end correct? A QSFP112 SR4 and a QSFP-DD DR4 will not link to each other regardless of what the cages do.

Vendors are often blunt about this in ways marketing copy is not. NVIDIA's interconnect documentation states plainly that QSFP112 is a single-port, four-channel 400G device intended for ConnectX-7 adapters and BlueField-3 DPUs, that QSFP112 ports are not for use in switches, and that single-port OSFP or QSFP112 devices cannot be used in twin-port OSFP switch cages at all - only in adapters and DPUs. That is the level of specificity you should be extracting from your own vendor's documentation before a purchase order goes out. Cisco publishes the equivalent information through its optics-to-device compatibility matrix, searchable by exact switch SKU.

Do not order against a sentence like "QSFP-DD is backward compatible." Order against a switch model, a software version, a port group and a module part number.

Side-by-Side Comparison

The Three Form Factors, Explained

QSFP-DD

Quad Small Form-factor Pluggable Double Density takes the four-lane QSFP interface and adds a second row of contacts, giving the host and module eight high-speed lanes. Those eight lanes have carried three generations: 400G on 50G-class lanes, 800G on 100G-class lanes, and 1.6T on 200G-class lanes in QSFP-DD1600, which the MSA introduced with hardware revision 7.0 in late 2023.

The commercial argument for QSFP-DD is continuity. A QSFP-DD cage can be designed to accept earlier QSFP modules, so a new switch can go in while selected QSFP28 or QSFP56 optics stay in service through a staged migration. That is a design intent, not a guarantee: the port still has to support the requested speed, the lane count, the management interface, the FEC mode, the power class and whatever qualification policy the vendor enforces. Our QSFP-DD technical overview goes through the electrical and management side in more depth.

OSFP

Octal Small Form-factor Pluggable also uses eight high-speed lanes, and the OSFP MSA defines the same 400G / 800G / 1.6T ladder as the lane rate climbs. What differs is the envelope: OSFP is physically larger than any QSFP variant, which buys module designers room for components and heat.

Standard OSFP can carry an integrated heatsink; the OSFP-RHS variant leaves cooling to a riding heatsink on the host. Up to 36 OSFP ports fit in a 1U faceplate.

OSFP-XD is a distinct thing and worth keeping straight. It doubles the electrical lanes from eight to sixteen, reaching 1.6T with 16×100G and pointing at 3.2T with 16×200G. It is not compatible with standard OSFP - keying features on the modules physically prevent insertion into the wrong port type. So OSFP-XD is not "the road to 1.6T"; standard OSFP already has one.

OSFP does not mate with QSFP-DD or QSFP112. A QSFP-family module cannot be pushed into a normal OSFP cage. Cross-form-factor links are routine - an OSFP switch port reaching QSFP112 endpoints over a validated cable or breakout - but that is a link-level arrangement, with each end keeping its own host form factor.

QSFP112

QSFP112 is a four-lane QSFP module built for roughly 112 Gb/s per lane, giving 400G aggregate from four lanes instead of eight. It stays inside the familiar QSFP mechanical family, which is exactly why it gets misidentified.

Its role is narrower than the other two. In NVIDIA's 400G LinkX documentation, QSFP112 is the adapter-side and DPU-side interface: ConnectX-7 ships in both OSFP and QSFP112 versions, while BlueField-3 platforms accept QSFP112 only. The switch side of those links is OSFP.

QSFP112 is also not an 800G part and will not become one. Four-lane 800G means 200G-class lanes, which is the QSFP224 generation - SNIA's SFF-8665 covers QSFP10, QSFP14, QSFP28, QSFP56, QSFP112 and QSFP224 as one family with very different electrical requirements.

Compatibility Has Four Separate Layers

Most bad purchases come from treating compatibility as one yes-or-no question. It is four, and all four have to pass.

1. Mechanical

Will the module enter and latch? A QSFP module can be accepted by a QSFP-DD cage designed for it; a QSFP112 module can be accepted by a QSFP-DD connector and cage; a QSFP module cannot enter an OSFP cage; an OSFP module cannot enter an OSFP-XD cage. That is the whole layer, and it is the least informative one.

2. Electrical lanes

Lane count, per-lane rate, PAM4 or NRZ, host SerDes capability, gearbox needs, and which port modes the platform exposes. A 400G QSFP-DD host on 8×50G and a 400G QSFP112 host on 4×100G move the same traffic over genuinely different electrical interfaces. The Ethernet side of this is specified in IEEE Std 802.3df-2024, which added the 400G and 800G physical layers and management parameters now used across all three form factors.

3. Firmware and management

The host must recognise the module identifier, apply the right management page, permit the module's maximum power, accept the vendor coding, agree on FEC, and support the breakout configuration you intend to use. A module that is mechanically and electrically fine can still be rejected outright, held in a warning state, or silently capped at a lower speed.

4. Optical and link

The form factor names the host interface. It says nothing about fiber, reach or connector. Those come from the PMD you select - SR4/SR8 for short multimode runs, DR4/DR8 for parallel single-mode, FR4/LR4 for duplex single-mode, coherent ZR/ZR+ for interconnect distances, or DAC/ACC/AEC/AOC for direct assemblies - and then from the plant itself: multimode or single-mode, MPO/MTP or duplex LC, polarity, distance, and insertion loss budget. Fiber grade matters here more than most quotations admit; the practical reach differences between OM3, OM4 and OM5 decide whether an SR-class link is viable at all.

Here is what those four layers produce in the combinations people actually ask about:

400G, 800G and 1.6T Roadmaps

Roadmaps describe specifications, not the switch on your quotation. Both eight-lane form factors reach 1.6T on 200G-class lanes - QSFP-DD1600 in the QSFP-DD MSA, OSFP1600 in the OSFP MSA. QSFP112 stays at 400G, with the four-lane family's 800G step belonging to QSFP224.

None of that changes what an installed switch can do. A 400G ASIC does not acquire 200G SerDes because you plugged in a 1.6T module. Keep four things apart when someone waves a roadmap at you: what the specification permits, what the ASIC in the chassis supports, what the specific port supports, and what the module you are buying actually is. The gap between the first and the third is where budgets die.

Power, Cooling and Port Density

There is no rule that QSFP112 modules draw less than QSFP-DD or OSFP modules. Power tracks the optics, not the shell: reach, laser technology, optical lane count, DSP or LPO architecture, coherent versus direct-detect, gearbox design, operating temperature, and the vendor's implementation. Comparing a short-reach parallel optic against a coherent ZR module by form factor is meaningless.

What the form factor does influence is headroom. OSFP's larger body gives designers space for components and heatsinking, which is why it hosts higher-power optics comfortably - but plenty of OSFP modules are modest. QSFP-DD's density is its own risk: pack a row of high-power modules side by side and the local heat load becomes a cage, airflow and inlet-temperature problem long before it becomes a module problem. QSFP112, with a flat top, leans on the fins built into the adapter cage.

Concrete numbers only exist per product. NVIDIA's MMS1X00-NS400, a 400G single-port QSFP112 single-mode transceiver for BlueField-3 and ConnectX-7, is documented at 9 W maximum when running 400G over four channels with MFP7E30 crossover fiber - and that figure applies to that part, in that mode, and nowhere else. Ask suppliers for the same specificity:

• maximum module power, not typical;
• the power class the host port is willing to supply;
• required airflow direction and rate;
• maximum case temperature and operating range;
• derating at high inlet temperature or altitude;
• whether every port can be populated at once at the stated power.

That last question is the one vendors answer least willingly and the one that most often bites.

Which Form Factor for Which Deployment

Upgrading an existing 100G or 200G data center

Evaluate QSFP-DD first. If the estate already runs QSFP28 or QSFP56 and the new switch supports those port modes, the migration can be staged: bring in the QSFP-DD switch, keep selected 100G optics running on the ports that allow it, use 400G-to-4×100G breakout where the platform permits, move server and leaf connections in batches, and convert ports to native 400G or 800G as budget allows.

Breakout is where this plan usually fails, and it fails structurally rather than optically. Many switches enable breakout only on selected ports or port groups, and some require a mode that reconfigures neighbouring ports. Confirm the port map before the fiber plant is designed, because the plant follows from it - a 400G DR4 port broken out to four 100G endpoints needs an MPO breakout assembly with the right fiber count and polarity, and getting that wrong is expensive to fix after installation.

Greenfield AI or HPC cluster

Follow the ecosystem instead of standardising on one shell. A current AI fabric typically puts OSFP on the switch, QSFP112 on the 400G NIC or DPU, and a validated OSFP-to-QSFP112 DAC, ACC, AOC or transceiver pair in between. NVIDIA's own 400G configuration is exactly this shape: Quantum-2 InfiniBand and Spectrum-4 Ethernet switches are OSFP-based, ConnectX-7 comes in OSFP or QSFP112, BlueField-3 is QSFP112 only, and a twin-port 2×400G OSFP transceiver in the switch can feed two QSFP112 endpoints over separate fibers.

The two ends of a link do not need matching shells. They need a validated end-to-end design, and the vendor's parts list is the authority on what counts as validated. Where the fabric is built on structured MPO trunking rather than point-to-point cables, the trunk cable choice - fiber count, polarity method, APC or UPC endface - has to be settled at the same time as the optics, not after.

400G server NIC or DPU

Start from the adapter's official cable and transceiver list and do not deviate. Check whether the NIC runs Ethernet, InfiniBand or both; what FEC it requires; which optical PMDs are listed; the exact cable part number; the firmware version; and whether dual-port or breakout behaviour is supported in your configuration. A generic 400G QSFP-DD module is not a substitute for a listed QSFP112 part just because both boxes say 400G.

400G or 800G data center interconnect

QSFP-DD and OSFP are both plausible. Decide on the host platform, coherent module availability, maximum module power, thermal design, required reach and the router vendor's qualification - not on a general claim about which shell is more efficient. For anything past the campus, the fiber grade drives the outcome as much as the optic does; the loss and dispersion differences between OS1 and OS2 single-mode fiber matter over ZR distances in a way they never do inside a row.

A Verification Sequence That Actually Works

The order below is not arbitrary. Each step invalidates the ones after it if it fails, so running them out of sequence wastes time.

1. Host model. The exact switch, NIC or DPU model - not the product family. Different models under one family name use OSFP, QSFP-DD or QSFP112.
2. Port mode. Lane count, per-lane rate, maximum aggregate, which slower modes are permitted, and which ports allow breakout. Read this from the datasheet, not from the front panel.
3. Firmware. The running software version, and whether the module identifier you intend to buy is in its supported list.
4. Module ID and coding. The exact part number and the vendor coding the host expects.
5. FEC. Confirmed as identical on both ends before the link is attempted.
6. Link test. Bring the port up and let it stabilise.
7. DDM check. Read optical power, temperature and error counters. A link that comes up with marginal receive power or a climbing pre-FEC error count has not passed - it has postponed failing until the room warms up.

Steps 5 through 7 are the ones most often skipped, and they are the ones that separate a link that works in a lab from a link that works in August.

FAQ

Can I plug a QSFP112 module into a QSFP-DD port?

Mechanically yes; operationally only if the host supports a four-lane 100G/112G-class port mode and the firmware admits the module. Check the platform documentation for that specific model.

Is QSFP-DD backward compatible with QSFP28?

That is the design intent of the form factor. Whether your port does it depends on the switch model, the port group and the firmware.

Is OSFP better than QSFP-DD?

Neither wins in the abstract. OSFP gives more thermal headroom and dominates AI/HPC switching; QSFP-DD gives migration flexibility and density. The host platform decides.

Which form factor uses the least power?

None of them. Power comes from the optics - PMD, reach, DSP architecture - not the shell.

Does QSFP112 support 800G?

No. Four-lane 800G is QSFP224, a different electrical generation.

Do both OSFP and QSFP-DD support 1.6T?

Yes, both MSAs define 1.6T on eight 200G-class lanes. Your existing switch still has to supply those lanes.

Can an OSFP switch connect to a QSFP112 NIC?

Yes, through a vendor-validated OSFP-to-QSFP112 cable, breakout or transceiver pair. Each end keeps its own form factor - nothing is being inserted across families.

Is OSFP-XD just a bigger OSFP?

No. It has sixteen electrical lanes instead of eight, and keying prevents cross-insertion with standard OSFP in either direction.

The Bottom Line

Choose QSFP-DD for migration flexibility inside a QSFP estate and access to 400G, 800G and QSFP-DD1600 platforms. Choose OSFP when the platform is OSFP-based, thermal headroom is a real constraint, or 800G and 1.6T are on the roadmap. Choose QSFP112 when a NIC, DPU or switch explicitly provides a four-lane 400G QSFP112 cage - and expect the other end of that link to be OSFP.

When you send a request to a module supplier, give them: equipment brand and exact model, operating system and firmware version, port type and requested speed, breakout mode, distance, fiber type, connector, operating temperature and preferred coding. Anything less and you are both guessing.