AI pushes silicon photonics integration

Nvidia has announced two networking switch product lines that combine electronics and silicon photonics in a single package to improve speed while reducing energy consumption. The Spectrum-X for Ethernet and the Quantum-X for Infiniband platforms offer speeds of 1.6 terabits per second per port to deliver a total bandwidth of up to 400 terabits per second at an energy consumption 3.5 times less than the currently available state of the art. Additionally, the new integrated products require 4x fewer lasers, offer 63x greater signal integrity and 10x better network resiliency at scale.

Nvidia says performance improvements like these are necessary to keep the AI revolution going. Currently, the speed at which data can be exchanged within a data center used to train AI models imposes a major bottleneck on raw computing power, despite utilizing optical connections. Additionally, the energy consumption of AI training is increasing exponentially. A more intimate collaboration between electronics performing the computations and photonics used for communication is coming to the rescue.