Back to search

Cornell University

Toward Fast Deployment and Resource Efficient AI on FPGAs

Abstract

dc:description.abstract

As machine learning models flourish, many hardware accelerators have been proposed, but designing an application specific accelerator for each new model or algorithm is not straightforward and usually requires multiple iterations between hardware and software. Field programmable gate arrays (FPGAs) have therefore become popular as platforms for prototyping and evaluating new architectures before fabrication. However, it is still unclear how to deploy machine learning models on these devices quickly while using logic, memory, and interconnect resources efficiently. Doing so often demands detailed knowledge of FPGA architecture and tool flows. This work addresses this gap through four representative case studies that span inference, training, emulation, and compilation, moving toward fast deployment and resource efficient AI on FPGAs. First, we design a compact object detection model and accelerator for nearby drone detection that fits on a small device through model simplification and quantization. Second, we propose a low-precision normalization algorithm and corresponding hardware design that preserve training accuracy close to full precision. Third, we introduce a chain-based time-division multiplexing (CTDM) method that emulates very large neural processing units on one or more FPGA devices while significantly reducing resource utilization. Finally, we present Torch2FPGA, a compilation framework that maps vision transformer models from PyTorch to a modular accelerator architecture, reducing manual design effort and shortening the path from new vision transformer models to running hardware prototypes.

Degree

thesis:*
Name thesis:degree_name
Ph. D., Electrical and Computer Engineering
Level thesis:degree_level
Doctor of Philosophy
Discipline thesis:degree_discipline
Electrical and Computer Engineering
Grantor
Cornell University
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Suh, Han-sok
Chair dc:contributor.chair
  • Seo, Jae-sun
Committee members dc:contributor.committeemember
  • Abdelfattah, Mohamed
  • Zhang, Zhiru

Subjects

dc:subject × 5

Rights

dc:rights
Statement dc:rights
  • Attribution 4.0 International
Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Bibid
17423409
Dc Identifier Other
ProQuest Submission ID: 15410
ProQuest Publication ID: 32396143

Chain of custody

source
Harvested from
Cornell University
Base URL
ecommons.cornell.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Suh, Han-sok. Toward Fast Deployment and Resource Efficient AI on FPGAs. Doctor of Philosophy thesis, Cornell University, 2025. https://hdl.handle.net/1813/121020