Copper Wires Have Already Failed Clustered AI Systems
The Next Platform, Tuesday, September 10th, 2024
It has become a well known fact these days that the switches that are used to interconnect distributed systems are not the most expensive part of that network, but rather it is the optical transceivers and fiber optic cables that comprise the bulk of the cost.
Because of this, and the fact that optical components run hot and fail often, people have shied away from using optics except where necessary.
And so we have copper cables, increasingly driven directly off switch ASICs and the devices they connect, for the short haul and optic cables for the long haul that is necessary to provide the 1,000 or 10,000 or 100,000 devices required for AI and HPC systems.