Back to search

Georgia Institute of Technology

COMPUTE-IN-MEMORY WITH EMERGING NON-VOLATILE MEMORIES FOR ACCELERATING DEEP NEURAL NETWORKS

Abstract

dc:description.abstract

The objective of this research is to accelerate deep neural networks (DNNs) with emerging non-volatile memories (eNVMs) based compute-in-memory (CIM) architecture. The research first focuses on the inference acceleration and proposes a resistive random access memory (RRAM) based CIM architecture. Two generations of RRAM testchips which monolithically integrate the RRAM memory array and CMOS peripheral circuits are designed and fabricated using Winbond 90 nm and TSMC 40 nm commercial embedded RRAM process respectively. The first generation of testchip named XNOR-RRAM is dedicated for binary neural networks (BNNs) and the second generation named Flex-RRAM features 1bit-to-8bit run-time configurable precision and leverages the input sparsity of the DNN model to improve the throughput and energy efficiency. However, the non-ideal characteristics of eNVM devices, especially when utilized as multi-level analog synaptic weights, may incur a notable accuracy degradation for both training and inference. This research develops a PyTorch based framework that incorporates the device characteristics into the DNN model to evaluate the impact of the eNVM nonidealities on training/inference accuracy. The results suggest that it is challenging to directly use eNVMs for in-situ training and resistance drift remains as a critical challenge to maintain a high inference accuracy. Furthermore, to overcome the challenges posed by the asymmetric conductance tuning behavior of typical eNVMs, which is found to be the most critical nonideality that prevents the model from achieving software equivalent training accuracy, this research proposes a novel 2-transistor-1-FeFET (ferroelectric field effect transistor) based synaptic weight cell that exploits hybrid precision for in situ training and inference, which achieves near-software classification accuracy on MNIST and CIFAR-10 dataset.

Degree

thesis:*
Level thesis:degree_level
Doctoral
Department dc:contributor.department
Electrical and Computer Engineering
Grantor dc:publisher
Georgia Institute of Technology
Year dc:date.issued
2020

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Sun, Xiaoyu
Advisor dc:contributor.advisor
  • Yu, Shimeng
Committee members dc:contributor.committeemember
  • Lim, Sung Kyu
  • Li, Shaolan
  • Raychowdhury, Arijit
  • Seo, Jae-sun

Subjects

dc:subject × 3

Rights

Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
http://hdl.handle.net/1853/64998

Chain of custody

source
Harvested from
Georgia Tech
Base URL
repository.gatech.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Sun, Xiaoyu. COMPUTE-IN-MEMORY WITH EMERGING NON-VOLATILE MEMORIES FOR ACCELERATING DEEP NEURAL NETWORKS. Doctoral thesis, Georgia Institute of Technology, 2020. http://hdl.handle.net/1853/64998