In the huge datacenters dedicated to artificial intelligence, light could prevent thousands of chips from wasting time waiting for one another.
To understand the problem, we need to look at how large AI systems work. A model can be distributed across many graphics processors, or GPUs, each of which computes part of the result. These chips must regularly exchange and recombine their data. The more of them there are, the more these communications can slow down the whole system.
Today, these exchanges mainly take place through electronic networks and switches. Researchers propose replacing part of this process with light, which can simultaneously transmit information to multiple locations.
This architecture is called Parallel Photonic Integration, or PPI. Here, light is involved in the calculations and exchanges that coordinate multiple machines.
One of the operations studied is AllReduce. Behind this name lies a frequent task during AI training: several GPUs produce partial results that must be combined and then redistributed. With many processors, this step can become lengthy. PPI could perform these combinations with light and in parallel.
The authors carried out an initial experimental demonstration of their architecture. They also used computer models to estimate its behavior in large GPU clusters running Transformer-type models. At very large scale, their calculations predict an AllReduce latency reduced by more than 100 times compared with that of current switched networks.
The benefits are not limited to speed. The same simulations predict energy consumption more than 50% lower per operation. A chip waiting for an exchange to finish continues to consume electricity. Reducing these waiting times could therefore lower the system’s overall power consumption.
These figures are still projections. In particular, the researchers will need to verify how this architecture performs when integrated into real infrastructures containing very large numbers of GPUs.