Communication Products

Optical Transceivers · Telecom Synchronization · Network Infrastructure · Custom Solutions

Home > Products > Optics Selection Guides > Data Center Leaf-Spine 400G: Optics Selection Guide
Optics Selection Guides

Data Center Leaf-Spine 400G: Optics Selection Guide


What a 400G leaf-spine fabric asks of its optics
A leaf-spine fabric is a two-layer network: every leaf switch connects to every spine switch, and every server or storage node connects to one or more

What a 400G leaf-spine fabric asks of its optics

A leaf-spine fabric is a two-layer network: every leaf switch connects to every spine switch, and every server or storage node connects to one or more leaves. The port rate is the same across the fabric, but the roles are not. Server facing ports run short links inside a rack or across a row, while leaf to spine links run longer and carry the oversubscribed aggregate of everything below them.

At 400G the two roles end up on different optics, and the reach classes tell them apart. This guide covers the standard positions in a 400G fabric, what decides the choice at each, and how the cabling carries into 800G. Rate families are listed on the 400G QSFP-DD and OSFP and 800G pages.

Server to leaf and leaf to spine are not the same link

Server to leaf links are usually under 100 m and often inside a single row, so multimode parallel optics, direct attach cables or active optical cables cover them at low cost per port. Where a row spans more than a hundred metres, an SMF parallel part at 500 m to 2 km replaces the multimode link.

Leaf to spine links run between rows, halls or down a building riser, and they carry the aggregate of many server ports. A 400G fabric commonly runs these on single-mode parallel or duplex WDM parts, chosen by the measured route, because a spine link that only needs 500 m is a different purchase from one that needs 10 km.

Recommended optics for a 400G fabric

PositionModuleMedia / wavelengthReach classWhere it usually lands
Server to leaf, in row400G multimode parallel, DAC or AOCMMF, 8 or 16 fibre MPO50 to 100 m classTop-of-rack links inside one row
Server to leaf, longer400G DR4 or DR4+SMF, 8-fibre MPO, 1310 nm500 m to 2 km classSingle-mode cabled rows and hall to hall
Server to leaf, duplex only400G FR4 or LR4SMF duplex, CWDM or LAN-WDM2 to 10 km classWhere no MPO trunk is available
Leaf to spine, short400G DR4 or DR4+SMF, 8-fibre MPO, 1310 nm500 m to 2 km classSpine links within one hall
Leaf to spine, campus400G FR4 or LR4SMF duplex, CWDM or LAN-WDM2 to 10 km classBetween buildings or down a riser
Leaf to spine, metro or campus backhaul400G ZR or ZR+ coherentSMF, DWDM C-band tunableMetro DCI class upwardWhere fibre is scarce or the span is tens of kilometres
Breakout at the leaf400G to 4 x 100GMPO trunk with 100G DR1 or PSM4 partsPer lane classOne spine port feeding four lower rate ports

Reach classes are the standard ones for each family, and the working limit is the measured loss through the real patch field. Parallel single-mode positions are covered by parts such as the APX-QSFP-DD-400G-2KM-DR4P, and the coherent end of the fabric by the 400G ZR DCI selection guide.

Cabling, MPO polarity and the strand maths

Single-mode parallel parts need eight-fibre MPO trunks, and the strand count grows four times faster than the port count. A row of forty 400G links needs three hundred and twenty strands on the parallel route, against eighty on the duplex route. The trunk is usually the long lead item, so the cable plan belongs in the planning phase.

Polarity is the other item that fails at turn-up: an MPO trunk wired for one polarity with optics that expect another shows a link that never comes up. Fix the polarity on paper, keep the patch record, and use one trunk type throughout the fabric.

Oversubscription, port counts and where the rate changes

The point of a two-layer fabric is predictable oversubscription: the spine carries the sum of the leaves. A leaf with thirty-two downlinks can be built with eight or sixteen uplinks and left with a two or four to one ratio, and the uplink count sets the spine port count. Compute the uplink count before the optics, because a plan that moves from four to one to two to one doubles the spine ports and the leaf-to-spine modules.

Where the rate changes inside the fabric, breakouts are the usual tool: a 400G spine port feeding four 100G leaf ports. Breakouts keep the trunk count down, but they also share the port between four services, so the failure domain belongs in the decision.

Temperature, monitoring and the 800G step

Fabric rooms run warm and the rows fill with high power modules. Confirm the temperature grade per position and the power per cage before the port count is fixed, and check that the module reports diagnostics to the monitoring platform.

The next step is usually 800G on the same trunk: the MPO or duplex pair already in place carries either, and the spine modules change rather than the cable. Builds that will migrate should fix the cabling class now; the worked comparison is on the 400G to 800G migration page, with the AI cluster case on the AI training cluster interconnect page.

How to choose, in order

  1. Split the link list into server facing and spine facing positions.
  2. Measure each type, including patching, and set the reach class from the measurement.
  3. Decide single mode or multimode per row, then the cable class and polarity.
  4. Fix the oversubscription ratio and the uplink count before the port count.
  5. Choose parallel or duplex at each position, and decide where breakouts are used.
  6. Set the temperature grade and the spare rule per part, and record the coding for the platform.

Frequently asked questions

Can a 400G fabric run on duplex fibre only?
Yes, with 400G FR4 or LR4 class parts at 2 to 10 km, or coherent parts where the span is longer. It uses fewer strands and costs more per port than a parallel part.

Should server links be multimode or single mode?
Decide by row length and by what is already cabled. Under about a hundred metres multimode is the cheaper position; single mode suits longer rows and a plant that is single mode throughout.

Is a DAC or an AOC an option at 400G?
Yes for short in-row links, and both avoid two optical ports. They are fixed length, so the rack layout has to be settled first.

When does a spine link need coherent optics?
When the span is tens of kilometres, fibre is scarce, or the link crosses sites. Inside a hall the grey and parallel parts cover it at lower cost.

What do you need to quote a 400G fabric?
The position list by role, the measured distances and loss, the cable class in place or planned, the oversubscription ratio, the platform per switch and the temperature at each end.

Send the position list, the measured routes, the cable class and the platform per switch, and we come back with a part per position, the reach class, the connector and polarity per link, and the spare lines for the fabric. Related reading: 800G DR8 against FR4 and ZR+, the 800G AI data centre interconnect guide, and all optics selection guides.