- This event has passed.
Peiyan Dong PhD Dissertation Defense
August 8, 2024 @ 3:00 pm - 4:00 pm
Name:
Peiyan Dong
Title:
Software-Hardware Co-Design: Towards Ultimate Efficiency in Deep Learning Acceleration
Date:
8/8/2024
Time:
3:00:00 PM
Committee Members:
Prof. Yanzhi Wang (Advisor)
Prof. David R. Kaeli
Prof. Devesh Tiwari
Prof. Cheng Tan
Abstract:
As AI techniques continue to advance, the efficient deployment of deep neural networks on resource-constrained devices becomes increasingly appealing yet challenging. Simultaneously, the proliferation of powerful AI technologies has raised significant concerns about sustainability and fairness, demanding increased attention from the community. This talk presents two novel software-hardware co-designs for improving the efficiency and sustainability of deep learning models. The first part introduces a hardware-efficient adaptive token pruning framework for Vision Transformers (ViTs) on embedded FPGA, HeatViT, which achieves significant speedup under similar model accuracy compared to the state-of-the-art. HeatViT is the first end-to-end accelerator for ViT on embedded FPGA and also achieve practical speedup by data-level compression for the first time. The second presents PackQViT and Agile-Quant, a paradigm of the efficient implementation for transformer-based models by sub-8-bit packed quantization and SIMD-based optimization for computing kernels. Our framework can achieve better task performance than state-of-the-art ViTs and LLMs with significant acceleration on edge processors, such as mobile CPU, Raspberry Pi and RISC-V. This work not only marks the first successful implementation of the LLM on the edge but also addresses the previous limitation where edge processors struggled to efficiently handle sub-8-bit computations. At the conclusion of the presentation, the speaker will discuss today’s challenges related to AI sustainability and fairness and outline her research plans aimed at addressing these issues.