Improving Deep Learning Application Throughput on Mobile Devices with Heterogeneous Processors
Year: 2024
Type: dissertation
Abstract: Improving Deep Learning Application Throughput on Mobile Devices with Heterogeneous Processors Choonghoon Park Department of Computer Science and Engineering College of Engineering The Graduate School Seoul National University The advancements in deep learning have led to the widespread adoption of deep learning applications. Consequently, mobile devices with CPUs, GPUs, and AI accelerators have been introduced. Numerous studies have focused on enhancing the throughput of deep learning applications on heterogeneous processors, but they targeted embedded boards or, in the case of mobile devices, did not consider accelerators. Therefore, this research proposes a method to improve throughput by applying quantization to FP32 networks and pipelining on mobile devices considering heterogeneous processors characteristics. To fully utilize the accelerator, weights and activations are quantized using post-training quantization. Subsequently, a genetic algorithm is used to explore mappings of pipeline stages to maximize the throughput. The effectiveness of this approach has been validated on a mobile device, Samsung Galaxy S22, achieving ×36 to ×43 throughput improve- ments for each network compared to single-processor mapping of FP32 networks. Keywords : Mobile Device, Heterogeneous Processors, Quantization, Pipelining Student Number : 2022-20126
Author: 박충훈
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7019876407
