Date Available

5-10-2023

Year of Publication

2023

Document Type

Doctoral Dissertation

Degree Name

Doctor of Philosophy (PhD)

College

Engineering

Department/School/Program

Electrical and Computer Engineering

Advisor

Dr. Ishan Thakkar

Co-Director of Graduate Studies

Dr. Salehi Ahmad

Abstract

Convolutional Neural Networks (CNNs) have proven to be highly effective in various fields related to Artificial Intelligence (AI) and Machine Learning (ML). However, the significant computational and memory requirements of CNNs make their processing highly compute and memory-intensive. In particular, the multiply-accumulate (MAC) operation, which is a fundamental building block of CNNs, requires enormous arithmetic operations. As the input dataset size increases, the traditional processor-centric von-Neumann computing architecture becomes ill-suited for CNN-based applications. This results in exponentially higher latency and energy costs, making the processing of CNNs highly challenging.

To overcome these challenges, researchers have explored the Processing-In Memory (PIM) technique, which involves placing the processing unit inside or near the memory unit. This approach reduces data migration length and utilizes the internal memory bandwidth at the memory chip level. However, developing a reliable PIM-based system with minimal hardware modifications and design complexity remains a significant challenge.

The proposed solution in the report suggests utilizing different memory technologies, such as Dynamic RAM (DRAM) and phase change memory (PCM), with Stochastic arithmetic and minimal add-on logic. Stochastic computing is a technique that uses random numbers to perform arithmetic operations instead of traditional binary representation. This technique reduces hardware requirements for CNN's arithmetic operations, making it possible to implement them with minimal add-on logic.

The report details the workflow for performing arithmetical operations used by CNNs, including MAC, activation, and floating-point functions. The proposed solution includes designs for scalable Stochastic Number Generator (SNG), DRAM CNN accelerator, non-volatile memory (NVM) class PCRAM-based CNN accelerator, and DRAM-based stochastic to binary conversion (StoB) for in-situ deep learning. These designs utilize stochastic computing to reduce the hardware requirements for CNN's arithmetic operations and enable energy and time-efficient processing of CNNs.

The report also identifies future research directions for the proposed designs, including in-situ PCRAM-based SNG, ODIN (A Bit-Parallel Stochastic Arithmetic Based Accelerator for In-Situ Neural Network Processing in Phase Change RAM), ATRIA (Bit-Parallel Stochastic Arithmetic Based Accelerator for In-DRAM CNN Processing), and AGNI (In-Situ, Iso-Latency Stochastic-to-Binary Number Conversion for In-DRAM Deep Learning), and presents initial findings for these ideas.

In summary, the proposed solution in the report offers a comprehensive approach to address the challenges of processing CNNs, and the proposed designs have the potential to improve the energy and time efficiency of CNNs significantly. Using Stochastic Computing and different memory technologies enables the development of reliable PIM-based systems with minimal hardware modifications and design complexity, providing a promising path for the future of CNN-based applications.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2023.191

Download

Share

COinS