Skip to Main Content
The current tendency in high performance computing systems is to increase the parallelism in processing at all levels utilizing multithreads, increasing the number of chips in racks and blades, as well as increasing the number of cores on a chip. The scaling of overall system performance that soon might approach Exaflop/s is, however, out of balance with respect to limited available bandwidth for shuttling ExaBytes of data across the system, between the racks, chips and cores. We will discuss how optics can provide huge power and bandwidth benefits closer and closer to the chip and eventually into the chip itself that would allow to overcome severe constraints of today's mostly copper I/O interconnects. Particular emphasis will be given to silicon photonics integrated circuits designed for low power off-chip and on-chip optical networks.