Files in this item



application/pdfFA20-ECE499-Thesis-Frank, Justin.pdf (1MB)
(no description provided)PDF


Title:Analysis of Cyclone V in computer vision applications
Author(s):Frank, Justin
Contributor(s):Cheng, Zuofu
Subject(s):Computer Vision
Cyclone V
Abstract:As embedded computing becomes more common in computer vision applications FPGAs have become a common solution to accelerate inference. Within Intel's line of FPGAs there exists ready documentation of the high cost and high power Arria and Stratix lines, but much less has been published on the performance of the lower cost and lower power Cyclone series devices. Data was collected for two popular frameworks: the semi-closed source OpenVINO and the open source PipeCNN project. Data was collected on inference time and power consumption for an array of popular models accelerated with OpenVINO across multiple CPU frequencies, multiple FPGA bitstreams, and multiple execution modes. For PipeCNN a design space exploration was carried out to get optimal performance and power numbers for a set of popular supported networks. For OpenVINO it was found that for most models heterogeneous inference outperformed CPU only inference. Further it was found that heterogeneous inference in general uses comparable power to CPU only inference. For PipeCNN it was found that performance had no strong tie to maximum utilization of any one resource on the FPGA. Overall these results show a compelling case for the use of Cyclone series FPGAs in embedded computing applications that require fast computer vision inference in relatively low cost and low power form factors.
Issue Date:2020-12
Date Available in IDEALS:2021-01-04

This item appears in the following Collection(s)

Item Statistics