It’s All About Energy: What it Takes to do a FLOP in Real Systems

Peter Hsu Visiting Scholar, EcoCloud Monday, November 19, 2018 @ 11:15 am BC 420 Hosted by: Prof. Babak Falsafi
Coffee, tea and croissants will be available before the talk as from 11:00 am

Abstract

Artificial Intelligence is based on computing, which consumes energy. As natural energy resources become more expensive and AI demands higher performance, the market naturally rewards computer architectures that deliver greater energy efficiencies. This talk looks at where energy goes in two real systems doing AI in very different environments: the IBM922 server, advertised on their website as “The best server enterprise AI,” and the Snapdragon 845 system-on-a-chip, used in the Samsung Galaxy S9 and many other Android mobile phones. I will walk through what it takes to do a single floating-point operation in each system, from getting the data in memory, physically through the package, board/substrate, on-chip wires, cache SRAM, register file, until finally reaching the ALU. Because the two systems were designed for very different markets and are at totally different scale, they achieve completely different levels of energy efficiency.

Biography

Peter Hsu received Ph.D. from University of Illinois Urbana-Champaign. He started at IBM T. J. Watson Research Center on the 801 Project. He joined SGI in 1990 as architect of MIPS R8000 TFP microprocessor, shipping 1994 in fifty TOP500 supercomputers. Peter co-founded ArtX 1997, developer of Nintendo GameCube. ArtX acquired by ATI Technology in 2000. After attending art school and learning to paint, Peter joined Oracle Labs 2011 as Architect on the RAPID project and built a 50K-thread parallel SQL database accelerator. Peter is currently visiting research at EPFL University in Switzerland and Consulting Computer Architect at Barcelona Supercomputing Center.