HNSE-P1-8. Deep Learning on a Xilinx Kria KV260 Vision AI Field-Programmable Gate Array Platform
Sarah Harris, Ph.D.1
Faculty Mentor: Venkatesan Muthukumar, Ph.D.1
1Howard R. Hughes College of Engineering, Department of Electrical and Computer Engineering
Deep Learning (DL) has revolutionized research and development over the past ten years. Several challenges are that DL requires large power consumption and can be slow. Field Programmable Gate Arrays (FPGAs) are great candidates for implementing DL algorithms and solutions because they are configurable and offer low latency and low power consumption. In addition, the FPGA platform has on-chip memory and computation accelerators, for example adders, which decrease memory bottlenecks and helps to prevent memory and bandwidth issues. Furthermore, due to the versatile architecture of the FPGA, users can design application-specific hardware, instead of using general-purpose hardware found in a processor. The Xilinx Kria KV260 Vision AI (KV260) FPGA board was used for this project as it contains numerous accelerated applications for performing DL with a live camera feed. Three major DL solutions were used: Facial Recognition (FR), Figure Detection (FD), and Object Identification (OI). Overall, FR and FD performed well, but OI currently demonstrated low accuracy. Each DL solution had ten test cases with FR, FD, and OI having 90%, 82%, and 57% respectively. These results show that the KV260 can successfully implement a variety of DL solutions, especially as an edge device. A future implementation is to improve upon the current DL training models to provide better accuracy.
Datino Dixon | Howard R. Hughes College of Engineering
Dr. Sarah Harris | Howard R. Hughes College of Engineering
Dr. Venkatesan Muthukumar | Howard R. Hughes College of Engineering