Committee: 

Dr. Esmaeilzadeh, Advisor     

Dr. Kim, Chair

Dr. Prvulovic

Dr. Chandra

Abstract:

The objective of the proposed research is to design a unified full-stack for accelerating deep neural networks (DNNs) using algorithmic insights. Using algorithmic insights promises to simultaneously provide orders of magnitude higher performance and energy efficiency over the conventional general purpose computing stack, while exposing high level abstractions to improve programmability. This thesis targets DNNs, an important machine learning workload that is both compute and memory intensive. Specifically, the proposed thesis aims to develop a specialized full-stack for accelerating DNNs that encompasses (1) a high-level algorithmic abstraction for the programming interface, (2) a specialized DNN accelerator architecture that uses algorithmic insights to gain orders of magnitude performance and efficiency over general purpose processors, and (3) a compiler that automatically and efficiently maps the operations in a DNN to accelerator architecture. The DNN full-stack is optimized for the constraints of the acceleration platform, which is the number of on-chip compute and memory resources for an FPGA platform, and the die area and power budgets for ASIC platform.