Tech, ExplainedNetworking, infrastructure and security
Part

Data centre interconnect (DCI)

AI training made these the busiest fibres.

INSIDE: OPTICAL TRANSPORT NETWORK (OTN)HOW IT WORKS · A 20-TERABYTE CHECKPOINT BETWEEN TWO DATA CENTRESAn AI trainingjob runs acrosstwo data centres40 km apartnear HyderabadEvery fewminutes thejob saves a20-terabytecheckpoint,which is160 terabitsDCI joins thebuildings with64 wavelengthsof 400 gigabitsa second eachTogether theycarry about25.6 terabitsa secondThe checkpointcrosses in about6 to 7 secondsThe light itselfneeds only about0.2 millisecondsfor the 40 kmALSO TALKS TOAI clusters andrack-scale systems↗ DATA CENTRESCross-connectsand interconnection↗ DATA CENTRESColocation↗ DATA CENTRESCo-packaged optics↗ AI AND DATA CENTRE CHIPSINSIDE: OPTICAL TRANSPORT NETWORK (OTN)HOW IT WORKS · A 20-TERABYTE CHECKPOINT BETWEENTWO DATA CENTRESAn AI training jobruns across twodata centres 40 kmapart near HyderabadEvery few minutesthe job saves a20-terabyte checkpoint,which is 160 terabitsDCI joins thebuildings with 64wavelengths of 400gigabits a second eachTogether they carry about25.6 terabits a secondThe checkpoint crosses inabout 6 to 7 secondsThe light itselfneeds only about0.2 millisecondsfor the 40 kmALSO TALKS TOAI clusters andrack-scale systems↗ DATA CENTRESCross-connectsand interconnection↗ DATA CENTRESColocation↗ DATA CENTRESCo-packaged optics↗ AI AND DATACENTRE CHIPSINSIDE: OPTICAL TRANSPORT NETWORK (OTN) · A20-TERABYTE CHECKPOINT BETWEEN TWO DATA CENTRESAn AI training job runs across two datacentres 40 km apart near Hyderabad↓Every few minutes the jobsaves a 20-terabyte checkpoint,which is 160 terabits↓DCI joins the buildings with 64wavelengths of 400 gigabits a second each↓Together they carry about25.6 terabits a second↓The checkpoint crossesin about 6 to 7 seconds↓The light itself needs only about 0.2milliseconds for the 40 kmALSO TALKS TOAI clusters and rack-scale systems↗ DATA CENTRESCross-connects and interconnection↗ DATA CENTRESColocation↗ DATA CENTRESCo-packaged optics↗ AI AND DATA CENTRE CHIPS
Data centre interconnect joins two data centres 40 kilometres apart with 64 wavelengths of 400 gigabits a second, so a 20-terabyte AI training checkpoint crosses between them in under 7 seconds.
opens a page inside this one a page in the same scene a page in another scene a concept page a label only, with no page

Data centre interconnect (DCI) is the fibre and optical equipment that connects one data centre to another, so that servers in separate buildings can exchange data at very high speed. Most DCI links are within one city or region, tens of kilometres long. Some cross a country. DCI is a special use of the optical transport network (OTN).

Why do data centres need to be linked at all? A large cloud provider rarely keeps everything in one building. It builds several buildings a few tens of kilometres apart, so that a fire or a power failure in one does not stop its customers' apps. Every piece of data written in one building is copied to another, often within milliseconds. The buildings must therefore be joined by links with enormous capacity and short delay.

AI training has made this demand grow faster than any other. An AI cluster can have tens of thousands of GPUs, the chips that train AI models. Often one building no longer has enough electricity for all of them. So the cluster is split across nearby buildings, and the GPUs in each building must exchange results with the others many times a minute. These are now among the busiest fibres anywhere. The AI accelerators page explains why these chips need so much power.

Consider a training job split between two data centres 40 kilometres apart near Hyderabad. Every few minutes, the job saves a checkpoint, a full copy of the model's current state, of 20 terabytes. The two buildings are joined by DCI that carries 64 wavelengths of 400 gigabits a second each, about 25 terabits a second in total. In bits, 20 terabytes is 160 terabits. So the checkpoint crosses to the second building in under 7 seconds. The light itself takes only about 0.2 milliseconds to travel the 40 kilometres.

Modern DCI is simpler than long-distance optical equipment. Short links need no amplifiers along the way. Often the optics come as small modules that plug straight into the data centre's switches, instead of separate cabinets. Inside each building, the link ends at cross-connects in a colocation hall or at the cluster's own network. Further in, co-packaged optics bring the same light right up to the switch chips.

DCI differs from the AI cluster networking inside a building. That network joins GPUs a few metres apart. DCI joins whole buildings, so delay and the cost of fibre matter much more.

Tech, Explained is a part of ingrid.fyi. Every page explains one piece of technology in plain words.