About the Event
Compute performance increased by orders of magnitude in the last few decades: Mega to Giga took about 15 years, Giga to Tera in 12 years, Tera to Peta in 11 years. This was made possible by continued technology scaling, improving transistor performance to increase frequency, increasing integration capacity to realize novel architectures, and reducing energy consumed per logic operation to keep power dissipation within limit. The technology treadmill will continue to fulfill the performance demand, and one would expect to reach Exascale level performance in about 10 years. However, it's the same Physics that helped you in the past will now pose some barriers: "Business as usual" will not be an option.
First, the energy and power will pose as a major challenge: an Exascale machine would consume in excess of a Giga-watt! Second, memory & communication bandwidth to feed such a machine with conventional technology would be prohibitive. Third, orders of magnitude increased parallelism, let alone explosion of parallelism created by energy saving techniques, would increase unreliability. And programming system will be posed with even sever challenge of harnessing the performance with concurrency.
This talk will discuss potential solutions in all disciplines, such as circuit design, architecture, system design, programming system, and resiliency to pave the road towards Exascale performance.
Shekhar Borkar graduated with M.Sc. in Physics from University of Bombay, MSEE from University of Notre Dame in 1981, and joined Intel Corporation. He is an Intel Fellow, Director of Exascale Technologies, and an IEEE Fellow. His research interests are high performance, low power digital circuits, and high-speed signaling.