Title:  Benchmark framework for 2-D/3-D integrated compute-in-memory based machine learning accelerator

Committee:

Dr. Shimeng Yu, ECE, Chair , Advisor

Dr. Sung-Kyu Lim, ECE

Dr. Saibal Mukhopadhyay, ECE

Dr. Muhannad Bakir, ECE

Dr. Michael Niemier, U of Notre Dame

Abstract: Neural-inspired compute-in-memory (CIM) accelerators with emerging non-volatile memory (eNVM) devices such as resistive random access memory (RRAM) have been proven in silicon for deep learning acceleration. We proposed an end-to-end benchmark framework for the software and hardware evaluation of CIM accelerators with versatile device technologies called DNN+NeuroSim. The proposed framework can support both inference and training chip evaluation, with wide range of technology parameters, from 130nm down to 7nm. Furthermore, as the 3-D integration was proposed as a promising solution to support high bandwidth and on-chip storage for machine learning platforms, we proposed the 3D+NeuroSim with extend 3-D featured parameters and integrated thermal model, to evaluate the monolithic and heterogeneous 3-D integrated CIM accelerators. We have done comprehensive benchmarks across different 2-D and 3-D integrated CIM accelerators and versatile device technologies to explore various design options, and released the proposed frameworks as public tools for research community.