I-Corps: Power Reduction for AI Computing

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

The broader impact/commercial potential of this I-Corps project is to advance artificial intelligence computing for a wide range of applications including in data centers and for unmanned autonomous vehicles. According to Forbes, in 2017 U.S. data centers consumed more than 90 Billion Kilowatt hours of energy. Due to Artificial Intelligence and Internet of Things, the expectation is that data center energy consumption will have grown by 2X by 2023, or 15% year-over-year increase. On the side of power-constrained machines, such as drones and robots, power consumption for computing is severely limited to milli-Watts of power. The impact to these industries, to enable AI inference computation is the power savings of 80X over traditional electronics. This I-Corps project develops ultra-low power reconfigurable nano-switches as memory. Artificial Intelligence inference cpmputing has high energy costs due to the large number of multiply-accumulate (MAC) functions required at each node in the neural network, often millions to hundreds of millions of operations for a single inference calculation. These ultra-low power reconfigurable switches store the AI model weights, and replace dynamic ram (DRAM) memory. Simulated results indicate 80X power reduction for each memory read needed for MAC operation. These switches are compatible with complementary metal-oxide semiconductor (CMOS) technology and can be produced at a minimal cost adder from a standard CMOS process. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.