Back to search

Cornell University

Trainable Fixed-Point Quantization for Deep Learning Acceleration on FPGAs

Abstract

dc:description.abstract

This thesis introduces a novel fixed-point quantization aware training tool, QFX, designed to bridge the gap between trained and deployed deep learning models on resource-constrained devices, such as embedded FPGAs. Conventional quantization techniques have primarily focused on quantizing only the matrix multiplication in deep learning models during training, necessitating extensive fine-tuning efforts to quantize other layers to fixed-point precision for FPGA deployments. In this thesis, QFX enables training “hardware-ready” deep learning models to address the problem, by effectively emulating the fixed-point casting function and basic arithmetic operations, and dynamically learning the binary-point position during model training. During deployment, the fixed-point operations in the model can be seamlessly transitioned to their synthesizable counterparts supported by HLS, which eliminates numerical issues. The performance is evaluated on image classification tasks based on accuracy on practical datasets. Moreover, the thesis introduces K-hot, a multiplier-free quantization strategy within QFX, designed to minimize DSP usage. The effectiveness of this strategy is demonstrated through integration with a state-of-the-art binarized neural network accelerator, showcasing improved hardware performance on an embedded FPGA. In summary, this thesis presents a tool that simplifies the quantization process, enabling the training of high-quality fixed-point quantized models with reduced quantization overhead. The tool is intended for open-source release, and can potentially expand the audience with less preliminary quantization knowledge needed in efficient machine learning. The innovative approach of applying multiplier-free quantization to binarized neural networks also implies future possibilities for deploying extremely compressed deep learning models on resource-constrained devices.

Degree

thesis:*
Name thesis:degree_name
M.S., Electrical and Computer Engineering
Level thesis:degree_level
Master of Science
Discipline thesis:degree_discipline
Electrical and Computer Engineering
Grantor
Cornell University
Year dc:date.issued
2024

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Dai, Dingyi
Chair dc:contributor.chair
  • Zhang, Zhiru
Committee member dc:contributor.committeemember
  • Abdelfattah, Mohamed

Subjects

dc:subject × 5

Rights

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

Identifiers

dc:identifier.*
Dc Identifier Bibid
16575437
Dc Identifier Other
ProQuest Submission ID: 11992
ProQuest Publication ID: 30692724

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

Dai, Dingyi. Trainable Fixed-Point Quantization for Deep Learning Acceleration on FPGAs. Master of Science thesis, Cornell University, 2024. https://hdl.handle.net/1813/115806