The future of optics in supercomputers

In the past decade rack-to-rack cables in the worlds largest computers have gone from being all electrical to all optical. The optimal strategy for tomorrow's supercomputers will be to exploit those technologies, such as optics, which continue to evolve. Several factors drive the spread of optics within the supercomputer. First, the supercomputer's network doesn't scale well, meaning that for constant optical bandwidth per node, the network's all-to-all bandwidth drops as the number of network nodes increases. Next, until we can integrate high speed optics right into the processor chips, bandwidth off the processor will be limited by the number of I/O cells. Since it will be hard to drastically increase the number of I/O cells on a chip of fixed size, we will have to run I/O cells faster, leading to the maximal electrical channel length becoming shorter, and leading to more I/O channels being converter to optics. This greater demand for optics in the system will be tempered by cost. Cost improvement in optics will not, however, come only from improvements in optics technology. It will also be determined by the evolution of optical standards and optical packaging. Optical packaging will evolve from the separable electrical-optical-electrical cable, to the active optical cable, to the all-optical router/switch chip, to the ultimate goal of making the optical connection at the processor chip. Many a supercomputer system is designed around the network and the connectors, and a 1st-level module to waveguide connector, coupled with an optical connector at the tailstock, should be the most effective combination for improving network performance of future systems.