Abstract

Multicore scaling—increasing the number of cores per die each generation—is currently the principal strategy of the microprocessor industry for continuing performance growth. As I will present in this talk, the results of our dark silicon study show that core count scaling will not provide the performance and value needed to scale down many more technology generations. This impending end of continued performance scaling and the looming end of Moore’s Law will be disruptive for the entire computing community. Significant departures from conventional approaches are needed to provide continued performance and efficiency gains in general-purpose computing. I will talk about general-purpose approximate computing as a new possible direction. I will then present a new class of accelerators, called Neural Processing Units. NPUs leverage an approximate algorithmic transformation that converts regions of code from a Von Neumann model to a neural model. Our work shows significant gains when the abstraction of full precision is relaxed in general-purpose computing and opens new venues for research.

Bio

Hadi Esmaeilzadeh is a Ph.D. candidate in the Department of Computer Science and Engineering at University of Washington. He has a master’s degree in Computer Science from The University of Texas at Austin and a master’s degree in Electrical and Computer Engineering from University of Tehran. Hadi is interested in developing new technologies and cross-stack solutions to improve the performance and energy efficiency of computer systems for emerging applications.