Description:

Recent advances in Machine Learning (ML) has resulted in widespread deployment of ML solutions in several areas. ML solutions are usually characterised by their high computational complexity. Heterogeneous computing platforms are introduced to optimise the performance and energy efficiency. They consist of a range of processing units, including Central Processing Unit (CPU), Graphics Processing Unit, Tensor Processing Unit (TPU), and Field Programmable Gate Array (FPGA). Each of the processing units on heterogenous platforms are sourced by its own specific voltage rail. This research will investigate the impact of undervolting on the accuracy of ML solutions implemented on AMD Xilinx heterogeneous platforms (Zynq MPSoC, Versal ACAP).

Type:

Undergraduate

Outcome:

• Investigating state-of-the-art ML solutions on heterogeneous computing platforms, including FPGA.
• Investigating the undervolting techniques on AMD Xilinx FPGAs and heterogeneous platforms. 
• Implementing a deep learning algorithm on FPGA as a case study and evaluating the impacts of undervolting on accuracy.
• Characterizing the voltage margins for the case study algorithm, and analysing the trade-off of energy saving versus accuracy loss.
• Developing a framework for power/energy analysis of ML solutions on AMD adaptive platforms.

Prerequisites

15 points from COMPSYS 305 

15 points from COMPSYS 304 or enrolment in COMPSYS 701 

Students should have a very strong background in digital systems design and have a good understanding of computer system architectures. Competency with Electronic Design Automation (EDA) tools is required. They should be willing to learn to work with new design tools from AMD.

Specialisations

• Computer Systems Engineering

Categories

• AI & Machine learning
• Embedded Systems

Supervisor

• Maryam Hemmati

Co-supervisor

Team

• Heath Grant
• Caleb Yang

Lab

Embedded Systems (405.760, Lab)