• Telecom & ICT Equipment Supplier
  • Multi-Vendor Sourcing
  • Worldwide Shipping
YT CIG Telecom & ICT Equipment
Get a Quote

OSFP vs QSFP-DD for 400G: form factor choice and port planning

Updated on 2026-11-19

OSFP and QSFP-DD both carry 400G over eight 50 Gb/s PAM4 lanes, so the electrical interface is the same. The cages differ: OSFP adds an integrated heatsink and accepts no QSFP28 modules, while QSFP-DD keeps the QSFP28 footprint and accepts 100G modules in the same port. That difference decides which cages a 400G port plan can reuse.

Cage mechanics and heatsink variants

OSFP stands for Octal Small Form Factor Pluggable, and revision 5.22 of the OSFP MSA specification, dated August 9, 2025, defines its mechanics. The module body is taller and wider than a QSFP-DD body because the standard OSFP module carries an integrated heatsink, either open-top or closed-top, with fins exposed to the rack airflow (OSFP MSA Rev 5.22, §3.3 and §3.4). OSFP-RHS removes those fins and makes flat contact with a riding heatsink that belongs to the host cage; both variants use the same connector and electrical interface, but mechanical keys stop one from seating in the other’s cage (OSFP MSA Rev 5.22, §1). The RHS variant exists because some chassis designers put the heatsink on the host and keep the module envelope smaller.

The QSFP-DD module keeps the QSFP28 cage footprint and doubles the high-speed lane count to eight, which lets a QSFP-DD port accept a QSFP28 or QSFP+ module at 100G or 40G (QSFP-DD MSA hardware specification). Its top is flat, and heat crosses the cage into the host heatsink; there are no fins on the module itself. At 400G the two cages reach the same faceplate count: a 1U Ethernet switch can mount 32 QSFP-DD cages, as the Juniper QFX5130-32CD does, and a 1U InfiniBand switch can mount 32 OSFP cages, as the NVIDIA QM9700 does (Juniper QFX5130-32CD datasheet; NVIDIA QM9700 user manual).

A QSFP-DD port is a backward-compatible cage, which matters when an existing 100G plant is reused: a QSFP28 LR4 module seats and runs at 100G. An OSFP cage has no such mode. OSFP, OSFP-RHS and OSFP-XD are the mechanical variants the MSA defines, and none accepts a QSFP28 or QSFP-DD module (OSFP MSA Rev 5.22, Table 1-1 and Appendix F).

OSFP and QSFP-DD at a glance

Parameter OSFP QSFP-DD
MSA document OSFP MSA Rev 5.22 (Aug 9, 2025) QSFP-DD MSA hardware specification
400G host interface 8 × 26.5625 GBd PAM4 (400GAUI-8; IEEE 802.3bs) 8 × 26.5625 GBd PAM4 (400GAUI-8; IEEE 802.3bs)
800G host interface 8 × 53.125 GBd PAM4 (OSFP800; IEEE 802.3ck) 8 × 53.125 GBd PAM4 (QSFP-DD800; IEEE 802.3ck)
Heat path Integrated heatsink, open or closed top; OSFP-RHS uses a host riding heatsink Flat module top; cage and host heatsink
Power reference 33 W ceiling for OSFP/OSFP800; 42.9 W for OSFP1600 (OSFP MSA Table 15-9) 12 W max for Arista QDD-400G-FR4; 16 W max for QDD-800G-2FR4 (Arista datasheets)
100G cage compatibility None defined QSFP28 and QSFP+ accepted
1U 400G switch example NVIDIA QM9700: 32 OSFP, 25.6 Tbps Juniper QFX5130-32CD: 32 QSFP-DD, 12.8/25.6 Tbps
Common role AI and InfiniBand fabrics, router line cards Ethernet leaf and spine switches

Compatibility answers

Can a QSFP-DD port take a 100G QSFP28 module? Yes. The cage keeps the QSFP28 footprint and defines backward operation with QSFP28 and QSFP+ modules (QSFP-DD MSA hardware specification), and the QFX5130-32CD also breaks its 400G ports out to 100/50/25/10GbE, up to 128 ports per 1U (Juniper QFX5130-32CD datasheet). Can an OSFP port take a QSFP-DD module? No: OSFP, OSFP-RHS and OSFP-XD are the only variants the OSFP MSA defines, and the cages use different keys and connector geometry (OSFP MSA Rev 5.22, Table 1-1 and Appendix F).

Do OSFP and QSFP-DD use the same optics? The optical engine can be identical, but the modules are not interchangeable. Arista’s 400G FR4 family is sold in OSFP and QSFP-DD housings with the same optical specification, and the part numbers and cages still differ (Arista QDD-400G-FR4 datasheet).

Electrical and thermal paths at 400G

Eight high-speed differential pairs reach the host in both cages, and 400G uses all eight. 400GAUI-8 uses 26.5625 GBd PAM4 signaling, which gives a 53.125 Gb/s line rate per lane; the interface carries 50 Gb/s of Ethernet payload per lane, for 400 Gb/s aggregate (OSFP MSA Rev 5.22, §13; IEEE 802.3bs-2017). The same eight-lane structure scales to 800G: OSFP800 and QSFP-DD800 run 53.125 GBd PAM4 per lane for 800 Gb/s, and IEEE 802.3ck-2022 adds 400GAUI-4 for four-100G-lane hosts and 800GAUI-8 for 800G. At the optical layer, 400GBASE-DR4 and FR4 put four 100G lanes on fiber, while 400GBASE-FR8, LR8 and SR8 use eight 50G lanes (IEEE 802.3bs-2017; IEEE 802.3cd-2018; IEEE 802.3cu-2021). The host always sees eight lanes, whether the module uses four or eight optical carriers.

Arista’s 400G FR4 family is built into both packages with the same optical engine, 2 km reach, four CWDM wavelengths and 12 W maximum (Arista QDD-400G-FR4 datasheet). OSFP modules draw 3.3 V, and the OSFP MSA defines power classes from 1.5 W in Class 1 to above 14 W in Class 8, with a 33 W ceiling for OSFP and OSFP800 (OSFP MSA Rev 5.22, Table 15-8 and Table 15-9).

A 12 W module gives neither package a thermal edge at 400G. Arista’s FR4 family is rated 0 to 70 C case temperature in both housings (Arista QDD-400G-FR4 datasheet), and the OSFP MSA calls that range its Standard class (OSFP MSA Rev 5.22, Table 11-1). The thermal path differs: OSFP fins sit in rack airflow and the cage needs only a cutout and retention, while QSFP-DD’s flat top makes the cage and host heatsink pull heat through a contact surface, so cage flatness, thermal interface material and airflow all affect the result. OSFP-RHS removes the fins and uses a host heatsink with a cage cutout (OSFP MSA Rev 5.22, §12).

Module power rises to 16 W at 800G in Arista’s QDD-800G-2FR4 (Arista QDD-800G-2FR4 datasheet), while the OSFP1600 class allows 42.9 W at 200G per lane (OSFP MSA Rev 5.22, §15.6 and Table 15-9), 2.7 times the 16 W module.

The 42.9 W ceiling matters for a 1.6T roadmap and matters little for a 400G fabric built on the modules above. The package with the higher ceiling has more surface area for heat rejection, which is why the OSFP body carries fins.

The NVIDIA QM9700’s 747 W typical draw, with passive cables, shows how much of the cooling job sits outside the module; its maximum with active cables is 1,720 W, and forward airflow is rated to 35 C with reverse airflow to 40 C (NVIDIA QM9700 user manual). That is about 23 W per port at the typical figure, counting switch silicon, fans and optics. The QM9700 needs front-to-back airflow, and the cage style changes the module interface while the rack’s cooling budget stays the same.

Power and cooling answers

What is the highest module power each MSA allows? The OSFP MSA publishes a 33 W ceiling for OSFP/OSFP800 and 42.9 W for OSFP1600 at 200G per lane (OSFP MSA Rev 5.22, Table 15-9). QSFP-DD power is managed through CMIS power classes, and current datasheet ratings are 12 W maximum for a 400GBASE-FR4 QSFP-DD module and 16 W for an 800G QSFP-DD 2FR4 (Arista QDD-400G-FR4 and QDD-800G-2FR4 datasheets).

Which package runs cooler at 400G? Neither has an edge at 400G, because the heat load and the 0 to 70 C case rating are the same in both housings (Arista QDD-400G-FR4 datasheet); the OSFP heatsink and higher power ceiling become relevant at 800G and above.

How many 400G ports fit in 1U?

Both packages reach 32 × 400G ports in 1U. The 32-port QSFP-DD switch moves 12.8 Tbps unidirectional or 25.6 Tbps bidirectional, and the 32-port OSFP switch moves 25.6 Tbps at 400 Gb/s per port (Juniper QFX5130-32CD datasheet; NVIDIA QM9700 user manual).

At 400G, the OSFP cage keeps the same 32-port 1U density as QSFP-DD. Its cages use two rows, and the 1U chassis is deep enough for the optics.

At 800G the arithmetic reverses: the Arista DCS-7280R4K-32DE-F puts 32 × 800G QSFP-DD ports in 2U and supports 64 × 400G or 256 × 100G at 25.6 Tbps (Arista 7280R4K datasheet). That is 64 logical 400G ports, twice the 1U 400G count.

Routing scale and packet buffer add a second density constraint. The Arista DCS-7280DR3A-36-F carries 36 × 400G QSFP-DD ports in 2U, 14.4 Tbps and a 16 GB packet buffer (Arista 7280DR3A datasheet), which is 18 ports per rack unit against 32 for a 1U 400G leaf. The 2U card trades faceplate density for port count per system and deeper buffering.

Rack height does not set the practical port count at 400G; a 10 kW cap does. The QM9700’s 747 W typical draw allows 13 switches, or 416 ports, before servers and cooling overhead are counted (NVIDIA QM9700 user manual). A 42U rack would hold 42 switches by height, or 1,344 ports, at more than 30 kW.

How should a 400G leaf allocate ports, breakouts and 800G headroom?

The brownfield Ethernet case belongs to QSFP-DD because a 400G QSFP-DD port accepts a 100G QSFP28 module. A 32-port switch can run 100G today and 400G after the server refresh without changing the chassis; on a duplex LC plant, the fiber also stays. The same 32-port switch breaks each 400G port out to 100, 50, 25 or 10GbE, up to 128 ports on the same 1U box (Juniper QFX5130-32CD datasheet). The QFX5130-32CD-AFO catalogue record lists the port map and the 1600 W redundant supplies, and the 100G QSFP28 family covers the optics that fill the same cages.

InfiniBand NDR is the clearest OSFP home: NDR switches use OSFP at 400G, and the NVIDIA ConnectX-7 400G adapter is a single-port OSFP card running NDR at 400 Gb/s (NVIDIA ConnectX-7 datasheet, OPN MCX75310AAS-NEAT). An Ethernet fabric built on OSFP switches can use the same 400G optics as a QSFP-DD fabric, but it cannot reuse installed QSFP28 stock. The NVIDIA ConnectX-7 400G record is the catalogue example of an OSFP host port.

Spares inventory has to follow the cage, because the OSFP and QSFP-DD versions are not interchangeable. Arista’s FR4 family is one example: both versions share the same optical specification, but the part numbers and cages differ, so a mixed fabric needs both SKUs. For a 400G Ethernet plan, buy QSFP-DD for leaf and spine switches that must absorb existing 100G QSFP28 optics, and buy OSFP when the switch, NIC or line card presents an OSFP cage. The 100G spine-leaf planning guide covers the mixed-rate port map.

Port and breakout arithmetic

Port planning starts with the budget for uplinks and downlinks, before any optic is chosen. A 32-port 400G switch has 32 slots, and every uplink slot is unavailable to a server or line card. Eight uplinks and 24 downlinks is the common 3:1 leaf: 8 × 400G = 3.2 Tbps up, 24 × 400G = 9.6 Tbps down, and 9.6 / 3.2 = 3:1. A 1:1 fabric uses sixteen uplinks and sixteen downlinks for 6.4 Tbps each way. Twenty-two downlinks with eight uplinks leaves two spare ports and 8.8 / 3.2 = 2.75:1.

A 36-port 2U switch such as the Arista DCS-7280DR3A-36-F can run 12 uplinks and 24 downlinks for 2:1, or 16 uplinks and 20 downlinks for 1.25:1 (Arista 7280DR3A datasheet). The same 24 downlinks broken out to 4 × 100G produce 96 × 100G server ports and the same 9.6 Tbps. Four spines with two 400G links per leaf consume all eight uplinks; each spine carries two links from every leaf, so eight leaves use 16 ports per spine, and a 32-port spine leaves 16 ports for WAN, storage or a second leaf group. Leave at least two ports per leaf unallocated when the fabric may grow, because adding a spine requires a link from every leaf.

400G PHY, reach and breakout options

PHY Standard Optical lanes Reach Connector Port breakout
400GBASE-SR8 IEEE 802.3cd-2018 8 × 50G 100 m OM4 MPO 2 × 200G or 8 × 50G
400GBASE-DR4 IEEE 802.3bs-2017 4 × 100G 500 m G.652 MPO-12 APC 4 × 100G
400GBASE-FR4 IEEE 802.3cu-2021 4 × 100G CWDM 2 km G.652 Duplex LC None over one pair
400GBASE-LR4 IEEE 802.3cu-2021 4 × 100G CWDM 10 km G.652 Duplex LC None over one pair
400GBASE-FR8 IEEE 802.3bs-2017 8 × 50G CWDM 2 km G.652 Duplex LC None over one pair
400GBASE-LR8 IEEE 802.3bs-2017 8 × 50G CWDM 10 km G.652 Duplex LC None over one pair

Breakout capability is the column that most often changes a port plan. 400GBASE-DR4 puts four 100G lanes on eight fibers and can split into four 100G links with an MPO-12 harness; the QDD-400G-DR4 record lists 500 m reach, MPO-12 APC and 4 × 100GBASE-DR interoperability. 400GBASE-FR4 and LR4 multiplex four 100G lanes onto one duplex pair, so a passive harness cannot separate them into four links. A 400G port fitted with FR4 carries one 400G link, while the same port fitted with DR4 can serve four 100G access ports.

800G headroom

Eight lanes remain at 800G, with the lane rate doubling to 53.125 GBd PAM4 for 800 Gb/s; IEEE 802.3ck-2022 defines 800GAUI-8 for that link (OSFP MSA Rev 5.22, §13; IEEE 802.3ck-2022). Arista’s QDD-800G-2FR4 presents two 400GBASE-FR4 interfaces, each with its own duplex LC connector, 2 km reach and a 16 W module maximum (Arista QDD-800G-2FR4 datasheet), so one cage yields two 400G links. That is how 32 × 800G cages become 64 × 400G logical ports on the Arista DCS-7280R4K-32DE-F (Arista 7280R4K datasheet). A QSFP-DD port accepts a QSFP-DD800 module later, and an OSFP port accepts an OSFP800 module later, provided the host supports the higher lane rate and thermal budget.

Neither cage accepts the other package, so the 400G purchase locks the 800G path. The 800G QSFP-DD 2FR4 record is the current catalogue example of an 800G module inside the QSFP-DD family.

Which optics and fiber plants match each port?

The fiber plant determines the optic and the port count. 400GBASE-DR4 carries four 100G lanes at 1310 nm over eight single-mode fibers, reaches 500 m on G.652 and terminates in MPO-12 APC; Arista’s QDD-400G-DR4 datasheet lists per-lane transmit power of -2.9 to +4.0 dBm and receive power of -5.9 to +4.0 dBm (Arista QDD-400G-DR4 datasheet; IEEE 802.3bs-2017). The APC polish has to match on the patching side. A duplex LC plant points to 400GBASE-FR4 or LR4: both put four 100G CWDM lanes on one pair, FR4 reaches 2 km and LR4 10 km over G.652 (IEEE 802.3cu-2021).

The QDD-400G-FR4 record lists 1271, 1291, 1311 and 1331 nm, with per-lane transmit power of -3.3 to +3.5 dBm and receive power of -7.3 to +3.5 dBm.

Those wavelengths sit on the 20 nm CWDM grid of ITU-T G.694.2, over the G.652 plant. 400GBASE-SR8 is the multimode option, with eight 50G lanes over OM4 to 100 m and an MPO connector (IEEE 802.3cd-2018); a building riser needs single-mode.

Two cabling rules keep the fiber plant consistent with the port plan. An MPO-12 DR4 plant and a duplex LC FR4 plant do not convert without a harness, and the harness changes the port count at the far end. The optic must also match the cage’s eight host lanes while using four or eight optical carriers; the CWDM and DWDM metro guide covers the wavelength-grid side when a 400G link leaves the data centre.

Frequently asked questions

Can one 1U switch mix OSFP and QSFP-DD ports?

Rarely, and only where the datasheet says so. Most 1U 400G platforms commit to one cage family across the front panel: the QFX5130-32CD is all QSFP-DD, and fixed-configuration OSFP switches are all OSFP. Mixed panels appear on some 2U and modular chassis, where each line card carries its own cage type. Check the platform datasheet before assuming either way.

Should I buy 800G-ready cages now?

Only when the platform datasheet lists the 800G module family for those cages. OSFP800 and QSFP-DD800 keep the cage geometry of their 400G parents, and the OSFP1600 ceiling of 42.9 W at 200G per lane (OSFP MSA Rev 5.22, Table 15-9) shows where the thermal headroom goes. A 400G-rated cage the vendor has not validated for 800G optics does not gain support through a later software release.

What plant does DR4 need compared with FR4?

DR4 runs 500 m over parallel single-mode on MPO-12 APC, eight fibers carrying four lanes; FR4 runs 2 km over duplex LC using four CWDM wavelengths near 1310 nm. The two plants are not interchangeable without a harness, and the harness re-count changes the port plan at the far end, so fix the optics family before the cabling order goes out.

When is OSFP-RHS the right pick?

When the rack is liquid-cooled or heat leaves through a cold plate rather than front-to-back fans. RHS, the riding-heat-sink variant, sits flat on the module top and lets the cold plate or chassis wall carry the heat, which the tall finned OSFP cannot do. Mainstream air-cooled switches do not use it.

How does a 400G port break out to 4×100G?

With an MPO-12 to four-duplex harness on DR4-style optics, or a breakout DAC for short runs. The switch must also support breakout configuration on that port group: the QFX5130-32CD documents up to 128 ports per 1U across its 32 cages (Juniper QFX5130-32CD datasheet). Breakout changes the effective port count in the leaf-spine math, so size the fabric against the broken-out numbers, not the cage count.

Need a specific part?

Send the part number and quantity — we confirm condition, availability and price before you commit.