| KEYWORD |
Study and Development of Reliable Accelerators for Vision Transformers
keywords AI ACCELERATOR, EMBEDDED SYSTEM, RELIABILITY, SPACE APPLICATION
Reference persons SARAH AZIMI, LUCA STERPONE
Research Groups DAUIN - GR-05 - ELECTRONIC CAD & RELIABILITY GROUP - CAD
Thesis type RESEARCH / EXPERIMENTAL
Description The advent of Vision Transformers (ViTs) has revolutionized computer vision tasks, showcasing remarkable performance on image classification and other vision-related applications. However, to deploy ViTs in real-world scenarios, addressing their computational demands becomes crucial. This proposal outlines a comprehensive study aimed at developing reliable accelerators for Vision Transformers execution.Leveraging systolic arrays as an efficient and high-performance hardware solution to accelerate Neural Network workloads, the research will commence with an open-source Systolic Array and concentrate on constructing a reliable and efficient ViT accelerator.
Objectives:
-Design and implement an HDL model for accelerating Vision Transformers using open-source hardware systolic array-based accelerators.
-The accelerator will be implemented on SRAM-based FPGA.
-Evaluate the impact of hardware faults on the performance of the accelerator and, consequently, on ViT execution at the application level.
-Investigate and propose mitigation strategies to enhance the reliability of the ViT accelerator in the presence of hardware faults.
Required skills Hardware Language such as VHDL or Verilog
Deadline 24/01/2025
PROPONI LA TUA CANDIDATURA