OSFP vs QSFP-DD at 800G: Platform Fit, Interoperability, and Deployment

Platform reference checked August 2026

NVIDIA ships Spectrum-4 in two boxes. The SN5600 presents 64 OSFP cages at 800G and 51.2 Tb/s. The SN5400 presents 64 QSFP-DD cages at 400GbE and 25.6 Tb/s. Same vendor, same silicon generation—different cage, and a different ceiling behind it. That is the shape of nearly every expensive 800G form-factor decision: not which family is better, but what the exact SKU in front of you presents, and what it will actually bring up.

This guide works at that layer—the module-to-host interface and the physical link. It covers the traps that survive a competent design review: cage variants that share an EEPROM identifier, connector conventions that a PMD name does not imply, breakout behaviour that changes the cabling BOM, and per-module support that varies inside a single platform family. It does not attempt to turn a cage choice into a theory of the entire AI network.

1. The Direct Answer: The Cage Does Not Tell You the Ceiling

The mechanical boundary is the part nobody argues about: an OSFP module will not enter a QSFP-DD port, and a QSFP-DD module will not enter an OSFP port. The boundary that actually costs money sits one level down—what the cage implies about speed, module support, and thermal class, and how often that implication is wrong.

The Spectrum-4 pair is the cleanest example. A QSFP-DD faceplate on an SN5400 is not an 800G faceplate, and the vendor name does not resolve it. The same holds in the other direction across the market: Cisco's Nexus 9364E-SG2 is sold as an OSFP model and a QSFP-DD model, and Juniper's QFX5240 family includes 800G OSFP and QSFP-DD variants—so a vendor-level rule of thumb fails in both directions at once.

2. What Changes Between OSFP and QSFP-DD—and What Does Not

Both families provide eight high-speed electrical lanes and have 800G implementations. The OSFP MSA maps that lane architecture from 400G through 800G and 1.6T. QSFP-DD adds a second row of contacts to the QSFP mechanical family; the QSFP-DD MSA emphasizes backward-compatible host cages that can accept earlier QSFP modules.

The same engine, two envelopes

At the level that matters for selection, both module families contain the same functional chain: host electrical interface, DSP/retiming and management, optical transmit and receive engines, and a fiber connector. The difference to design around is the envelope around that engine—card edge, heatsink arrangement, airflow path, and host cage.

Three thermal implementations for 800G pluggables. The host documentation—not a generic OSFP label, and not the EEPROM identifier—determines whether an IHS or RHS module belongs in the port.

Drawn side by side, the point is easier to see than to argue. The two block diagrams below use a matched pair of 800G DR8 modules—VO-8CDR8CM-AA in OSFP and VD-8CDR8CM-AA in QSFP-DD—so the only variables between them are the envelope, the cooling path, and the card edge.

Inside the OSFP envelope: one PAM4 DSP, eight driver-and-laser pairs, eight PIN receivers and a CMIS 5.0 MCU behind a 60-pad single-row card edge—with the heatsink carried on the module itself. The twin-port layout gives the cage a native 2×400G personality without a splitter panel.

The same engine in a tighter envelope. QSFP-DD carries an identical DSP-and-optics chain, moves the heatsink onto the cage, and spends the reclaimed space on a second contact row—76 pads against QSFP's 38—which is what buys backward compatibility.

3. Current 800G Platform Reality: The Same Network Role Can Use Either Cage

The table below is deliberately platform-specific. It is not a compatibility matrix for every optic, and it should not be generalized to every product from a vendor. Its purpose is to show why vendor-, protocol-, and workload-level shortcuts fail.

Reference checked August 2026. Always confirm the exact switch, line card or adapter SKU, software release, port mode, and vendor-supported optics list before deployment. Support also varies per module inside one family—Juniper, for example, does not list 800G SR8 support across its current 800G platforms.

Example placements in an 800G estate—CX-7 and ConnectX-8 adapter cages at the NIC edge, finned OSFP-IHS at leaf and spine, and a QSFP island alongside for storage, management and brownfield Ethernet. These are deployment examples, not form-factor rules: the exact adapter, switch and line-card SKU decides the cage.

4. Thermals and Mechanics: Compare Systems, Not Slogans

OSFP is physically larger and was designed with substantial thermal headroom. The current OSFP specification defines both integrated-heatsink and riding-heatsink systems. QSFP-DD achieves a smaller QSFP-family envelope and also supports host-side heatsink designs. Both are used for real 800G products.

A generic statement such as "OSFP runs cooler" is incomplete without the module power, heatsink, airflow, inlet temperature, port density, and host design. Juniper's current 800G guide places typical 800G pluggables in the 16–18 W range, with coherent ZR/ZR+ modules reaching higher. At 32 or 64 ports, optics become a material part of the switch and rack thermal budget.

1. Confirm the host's power allowance per port. Do not infer it from the form factor's theoretical power class.
2. Match IHS versus RHS exactly—and do not expect the module to tell you. Finned-top and flat-top OSFP modules both report EEPROM identifier 19h; the identifier separates OSFP from QSFP-DD (18h), not one heatsink variant from the other. The required cage type comes from the host. Our IHS vs RHS selection guide and the EEPROM guide cover what the fields do and do not disambiguate.
3. Check airflow and inlet limits with the intended optic. A port qualified with DACs is not evidence that a full faceplate of higher-power optics has the same thermal margin.
4. Budget the whole faceplate. Multiply real maximum module power by populated ports, then include system fans, switch ASIC, ambient conditions, and any vendor derating rules.
5. Keep the spare mechanically exact. "800G OSFP" is not a sufficient spare description when IHS/RHS, connector, reach, and firmware coding can differ.

DSP, LRO and LPO: the third variable inside the cage

The heatsink variant is not the only thing that changes a module's thermal class. The same cage—OSFP or QSFP-DD, integrated or riding heatsink—can carry three different signal architectures, and they do not share a power budget or a host requirement.

Power figures are indicative, not datasheet values. The exact module's datasheet and the host's per-port allowance govern.

This belongs in a form-factor guide because LPO changes two of the three gates. The thermal budget drops sharply—an LPO faceplate carries the same port count at roughly half the optics power—and the per-module support question moves from the platform's optics list into its silicon and firmware. A host that runs a DSP module in a given cage cannot necessarily run an LPO module in the same cage: the switch or NIC SerDes has to provide the equalization the module no longer does, and the software has to identify a linear module and manage it. The standards stack now exists—the LPO MSA completed its 100G-per-lane single-mode specification in 2025 (100G-DR-LPO, covering 100G through 800G parallel single-mode links to 500 m, with a 400G-FR4-LPO specification alongside it), OIF CEI-112G-Linear-PAM4 defines the host-to-module electrical interface, and the CMIS Versatile Control Set carries the management. Support is still listed per platform and per module, exactly as for DSP optics, only with a shorter list.

Two consequences for the design review. First, an LPO link has no DSP at either end to absorb a marginal channel, so host, module, fiber and far-end host are validated as one system; a fiber that passed with DSP modules is not evidence that it will pass linear. Second, DSP and LPO modules are not interchangeable spares even when they share a cage, a connector and a reach, so the spare record carries the signal architecture alongside the heatsink variant. LPO fits the short-reach scale-out tier—DR- and FR-class links inside the hall, where the power and latency savings are largest; longer reaches and coherent DCI stay with DSP. Our LPO guide covers the selection question in depth.

5. Can OSFP and QSFP-DD Operate on the Same Fiber Link?

Yes—at opposite ends of the link—provided the optical media type and link parameters match. Juniper's 800G FAQ explicitly notes that OSFP and QSFP-DD can interoperate on the same link when the Ethernet media type is the same. That is optical interoperability, not mechanical compatibility—and it is the property that makes a mixed estate survivable.

Coherent DCI makes the separation especially clear. Cisco offers the same 800G ZR/ZR+ optical application in both OSFP and QSFP-DD. The host form factor changes; the link job—an interoperable 800G wavelength over a defined optical path—does not.

The minimum link record

6. Breakout, Connector, and Fiber Choices: Start with the Port Mode

Breakout is not an OSFP-only capability. Cisco documents 800G OSFP modules with 800GE, 2×400GE, 4×200GE and 8×100GE options, while the Nexus 9364E-SG2 family supports breakout in both cage variants. Exact modes depend on the platform, software release, optic or cable, and the switch's lane-mapping rules.

Use this order:

1. Define the link job. Native 800G switch-to-switch, 2×400G server attachment, 4×200G fan-out, or migration into an existing lower-speed tier.
2. Confirm the switch port mode. Verify the mode on the exact platform and release before choosing the optic.
3. Select the optical interface. Match reach, fiber type, loss budget, lane count and operational environment.
4. Select the connector implementation. DR8 does not automatically mean dual MPO-12; 800G products can use MPO-16, dual MPO-12, duplex LC and other interfaces. This is the most common source of a breakout order that cannot be cabled as drawn.
5. Draw the lane map end to end. Include switch logical ports, module lanes, trunks, cassettes or fan-out assemblies, polarity, and the far-end port.

Breakout is not an OSFP privilege. A twin-port OSFP cage behaves as two transceivers—native 2×400G on two MPO-12 APC trunks, no splitter panel—while QSFP-DD800 runs straight 800G or fans into the 400G/100G estate it already lives in. Confirm the modes your platform and software release actually support.

Reach picks the media, media picks the fiber, and the module picks the connector. The Vitex DR8 and VR8 SKUs shown here terminate dual MPO-12 APC—one twelve-fibre ferrule per four-lane engine—so a single module runs one 800G link or two 400G links without a breakout cable. MPO-16 and duplex LC are separate conventions; match the module's datasheet and the plant you already own.

For detailed lane maps, see the 800G breakout configuration guide. For trunk design and polarity, use the 800G structured cabling guide.

7. Four Real Estate Patterns—and the Fiber Consequence of Each

This is as far as workload and architecture need to enter a form-factor guide. The broader application, traffic and neocloud story belongs in a separate workload-to-fiber pillar. Here, the useful question is how the physical estate changes the optic BOM.

Rail-optimized fat-tree, Spectrum-X Ethernet and InfiniBand NDR/XDR alter the wiring pattern and the breakout posture—none of them creates a universal form-factor rule. Ethernet AI can use OSFP or QSFP112; verify the exact platform generation and SKU.

Where the tiers sit: scale-up, scale-out, scale-across

AI fabrics are now described in three tiers, and the reach ladder in the wiring reference above maps onto them directly. Scale-up is the accelerator-to-accelerator domain inside a rack or pod: a few meters, copper today, where pluggable optics rarely appear. Scale-out is the fabric inside the hall—server to leaf, leaf to spine—from a few meters to a few hundred meters, and this is where every 800G pluggable decision in this guide lives: VR and SR on multimode in the row, DR on parallel single-mode to 500 m, FR-class to 2 km. Scale-across joins halls and sites: LR-class links at 10 km, then coherent ZR and ZR+ for DCI at 120 km, in either cage. Naming the tier first shortens the review, because the tier narrows the reach class, and the reach class narrows the PMD and the media before any form-factor question is asked.

The operating consequence

For an engineer, a weak form-factor assumption can produce a link that never comes up. At fleet scale it also fragments spares, labeling, qualification records and replacement procedures. For a data-center or commercial leader, that becomes delayed sellable capacity and a narrower second-source strategy. A clean port-level record is therefore both an engineering control and an operating asset—it is what makes a second source qualifiable instead of a gamble.

8. The 800G Form-Factor Design Review

A useful review ends with a port-level record, not a winner between two acronyms. Run the sequence below for every platform role in the design.

A dated, per-SKU snapshot rather than a vendor rule. NVIDIA's 800G switching and Arista's 7060X6 flagship present OSFP, and QSFP-DD800 is common across Cisco- and Juniper-class platforms—though several of those families ship in both cages, Cisco's Nexus 9364E-SG2 and Juniper's QFX5240 among them, as the table in section 3 shows. Note the SN5400 row: QSFP-DD cages with a 400GbE ceiling. Always check the platform's own dated supported-optics list before raising the purchase order.

Start from the exact estate: read the cage off the faceplate, or off the exact greenfield or brownfield SKU. Then clear three gates—per-module platform support from the dated supported-optics list, rack thermal budget against module power class and the exact host limit, and the full link match of port mode, PMD, connector, polarity, FEC and far end—before the purchase order goes out.

9. What Changes at 1.6T and with Co-Packaged Optics?

The next NVIDIA generation makes this guide's rule sharper, not obsolete. Spectrum-6 is built on 200G-per-lane SerDes: NVIDIA's Spectrum-X Photonics configurations present 128 ports of 800 Gb/s or 512 ports of 200 Gb/s at roughly 100 Tb/s, and 512 ports of 800 Gb/s or 2,048 ports of 200 Gb/s at roughly 400 Tb/s. An 800G port on that host is therefore four lanes at 200G, not the eight lanes at 100G that Spectrum-4 presents today. The cage can look the same; the lane rate behind it does not. Whether a given 8×100G module is supported in a 200G-per-lane cage at a reduced per-lane rate is a per-SKU supported-optics question—exactly the check this guide already asks at 800G.

Spectrum-6 ships both ways. The co-packaged Spectrum-X Ethernet Photonics switches put the optical engine on the switch package; the same SN6000 series also ships pluggable-optics variants—the SN6600-LD, for example, is a liquid-cooled 2U switch with 64 OSFP cages presenting 128 ports of up to 800 Gb/s on the existing pluggable-transceiver ecosystem, each cage exposed as two 800G links. Co-packaged optics therefore changes where the optical engine sits on the switch side of a link. It does not remove the pluggable from the fabric: ConnectX-9 supports an RHS cage only in its OSFP form factor and is documented at up to 800 Gb/s per port, so on a co-packaged switch fabric the link becomes asymmetric—a pluggable module at the server, an on-package engine at the switch—and the pluggable decision moves to the NIC rather than disappearing. Two consequences for the design review follow. The OSFP-RHS requirement at the NVIDIA NIC edge carries into the next generation, and QSFP-DD still has no NVIDIA host at 800G or above.

For the architecture, external-laser and serviceability details, see Vitex's complete CPO guide. The practical rule is unchanged: qualify the host and link that will actually be deployed, and re-open the record when the platform generation changes—the lane rate is now part of that record.

FAQ: OSFP vs QSFP-DD at 800G