Sprinting Toward the Practical Limits of Computation

Stan Williams, HP Labs

There have been many incorrect predictions about the end of computational scaling, but this will not be one of them. It is within the realm of fundamental physics to improve computation dramatically beyond where it is now (petaflops), but if a zettaflop system ever exists, it is surely many decades in the future. We can only count on strict Moore's Law scaling to improve computational efficiency by a factor of 100 or so, and even then at the cost of having thousands of cores on a single chip. Replacing electronic interconnect by photonic interconnect at all length scales over a few microns might improve computational efficiency by another factor of 100 over the long term.

However, getting that last factor of 100 will require a large number of tricks and technologies for which we have only an inkling now. This talk will provide a brief overview of some of the challenges and the tricks for building future ultra-scale computers.