Abstract

The rapid development and growing interest in Artificial Intelligence (AI) have resulted in a trend of models increasing in size and complexity at an accelerating rate. These larger models have demonstrated a greater ability to accomplish increasingly sophisticated tasks compared to smaller models, furthering the trend of making models larger. However, as the models continue to improve and grow larger, this has created a new challenge in deploying them into practical applications. Current models typically run inference on the same systems where they were trained. Large data centers comprising Central Processing Units (CPUs) and Graphics Processing Units (GPUs) are well-suited for running these large models. However, many of the envisioned applications would require a significant reduction in size and power utilization to be effectively implemented. This research will investigate the potential of a new number system called Posits and its ability to optimize models through quantization. The potential will be tested by quantizing two CNN networks from 32-bit floating-point to 6-bit Posits, culminating in the design and verification of a hardware accelerator that implements both models in Posit form.

Library of Congress Subject Headings

Neural networks (Computer science)--Energy consumption; Convolutions (Mathematics); Computer vision; Sound--Classification; Optical pattern recognition

Publication Date

8-2025

Document Type

Thesis

Student Type

Graduate

Degree Name

Computer Engineering (MS)

Department, Program, or Center

Computer Engineering

College

Kate Gleason College of Engineering

Advisor

Mark A. Indovina

Advisor/Committee Member

Dorin Patru

Advisor/Committee Member

Cory Merkel

Campus

RIT – Main Campus

Plan Codes

CMPE-MS

Download

Share

COinS