What AI data centre interconnect has to carry
An AI data centre is not one building with one fabric. Training clusters are built out in halls, a campus grows around them, and the links between halls, buildings and nearby sites carry the traffic that keeps the cluster working as one machine: checkpointing, storage replication, model distribution and east-west traffic between pods.
That mix sets the requirement. Ports in the back-end fabric have moved to 400G and 800G per port, hall-to-hall links need kilometres of reach without coherent cost, and campus or metro interconnect between sites is where coherent optics come in. The selection question is therefore about distance tiers first, then form factor, then fibre and cooling.
Why 800G per port
800G on one port raises the number of switch ports a fabric can use for the same rack unit count, which is what matters when a cluster needs a large number of high-rate endpoints in a small cabling distance. QSFP-DD and OSFP both carry 800G; the choice between them is usually decided by the switch side and by whether the port has to break out later.
Recommended optics for 800G AI interconnect
| Module | Rate | Media / wavelength | Reach | Typical use |
|---|
| 800G OSFP / QSFP-DD DR8 | 800G, 8 x 100G | SMF, parallel 1310 nm | 500 m class | In-row and same-hall links between switches in the back-end fabric |
| 800G OSFP / QSFP-DD 2xFR4 | 800G, 2 x 400G | SMF, CWDM or LAN-WDM | 2 km class | Hall-to-hall links where a duplex pair is available |
| 800G OSFP 2xSR4 / SR8 | 800G | MMF, 850 nm parallel | 50 to 100 m | Short in-row links where structured MMF already exists |
| 800G to 2 x 400G breakout | 800G host, 2 x 400G links | SMF or MMF plus breakout assembly | Span dependent | Feeding two 400G ports from one 800G cage |
| 400G QSFP-DD DR4 / FR4 | 400G | SMF, parallel or CWDM | 500 m to 2 km | Fabric tiers still running at 400G per port |
| 400G / 800G ZR and ZR+ coherent | 400G or 800G coherent | DWDM C-band, tunable | Campus to metro spans | Building-to-building and site-to-site links without grey reach |
| DAC and AEC assemblies | 400G / 800G | Copper, passive or active | A few metres | Adjacent racks where copper is sufficient |
The reach column lists the standard classes behind each code; the working limit is the cabling loss and the connector count on the actual route. The 800G form factors and reach codes are set out on the 800G transceiver page, the tier below on the 400G QSFP-DD and OSFP page, and the coherent side on coherent optics.
Distance tiers: in-row, hall-to-hall, campus, metro
Treat the interconnect as tiers. In-row and same-hall links are hundreds of metres at most and are served by parallel single-mode or multimode parts, DR8 or SR8 types, where the cabling is short and the port count is high. Hall-to-hall and building-to-building links move to the 2 km class, usually 2xFR4 optics on a duplex pair or a parallel single-mode part. Campus and metro links between sites are where coherent plugs in: a 400G or 800G ZR or ZR+ module on the DWDM grid carries the span without a transponder shelf at each end. State the tier for every link in the request, since the reach code and the fibre type follow from it.
Structured cabling, MPO and polarity
An 800G parallel port normally terminates on an MPO or MTP assembly, and that is where build-outs go wrong. Trunk polarity and the pin-out of each cassette have to match the transmit-to-receive mapping of the port. Order the optics and the cabling assemblies as one set, keep the polarity scheme consistent across the hall, and check the link length against the reach class after the patch panels and cassettes are counted, not before.
Power, cooling and panel density
An 800G module dissipates well above a 400G module in the same cage family, and the back-end fabric concentrates many of them per rack unit. Three checks belong before the count is fixed: the port power budget of the line card, the airflow available in the rack, and whether coherent modules are being mixed into the same shelf. Where thermal margin is tight, spreading high-rate ports across several line cards is more forgiving than concentrating them on one.
Coding, rate negotiation and spares
The module has to be supported by the host port and coded for the platform you name, with optical power, module temperature and alarms readable through the host. Modules supplied for branded switches are compatible parts and are not manufactured by the brand of the host, with no claim of relationship to that brand. Name the platform and firmware release, and say whether the far end will be a different vendor, since that decides the coding and the mode set. Platforms common in these fabrics include Mellanox, Arista, Cisco and Juniper platforms. Keep spares per link type: a failed optic in a training fabric costs cluster time, not one link.
How to choose, in order
- List the tiers in the build: in-row, same hall, hall-to-hall, campus, metro.
- Measure the route after patch panels: length plus connector and cassette loss.
- Set the rate per tier, 400G or 800G, from the switch port and the port count needed.
- Choose the form factor the host supports, QSFP-DD or OSFP, and confirm breakout plans.
- Pick the reach code and fibre type for each tier, and order cabling with the optics as one set.
- Check port power, airflow and panel density, then name the platform for coding and spares.
Frequently asked questions
Is 800G needed for AI interconnect, or is 400G enough?
It depends on the tier and the switch. 400G remains the working rate one tier down; 800G is chosen where panel density and port count per rack unit are the limit, which is typical at the top of a back-end fabric.
QSFP-DD or OSFP for an 800G link?
Both carry 800G. The switch port decides, and the second question is whether the cage has to break out to two 400G links later.
When do I need coherent optics between buildings?
When the span is longer than the grey reach class of the tier, or when fibre is scarce enough that several links have to share one pair on the DWDM grid. A 400G or 800G ZR or ZR+ module then replaces the transponder shelf.
Can existing multimode cabling be reused at 800G?
Only for the short in-row class where an SR-type parallel module is specified and the installed fibre grade supports the rate and distance. Hall-to-hall and longer links are normally single-mode.
What has to be in the request for an interconnect build?
The tier and measured route length, the switch platform and firmware, the port type and breakout plan, the fibre and cabling assemblies available, and the temperature and power limits of the cabinet.
Send the switch platform, the tier and measured distance, the port type and the cabling available, and we confirm the reach code, the form factor and the coding for each link. Related reading: 800G DCO transponder card where capacity has to be aggregated onto wavelengths, 400G transponder and muxponder cards for shelf-based transport, and all optics selection guides.
