Comparing Energy Efficiency of CPU, GPU and FPGA Implementations for Vision Kernels

Abstract

Developing high performance embedded vision applications requires balancing run-time performance with energy constraints. Given the mix of hardware accelerators that exist for embedded computer vision (e.g. multi-core CPUs, GPUs, and FPGAs), and their associated vendor optimized vision libraries, it becomes a challenge for developers to navigate this fragmented solution space. To aid with determining which embedded platform is most suitable for their application, we conduct a comprehensive benchmark of the run-time performance and energy efficiency of a wide range of vision kernels. We discuss rationales for why a given underlying hardware architecture innately performs well or poorly based on the characteristics of a range of vision kernel categories. Specifically, our study is performed for three commonly used HW accelerators for embedded vision applications: ARM57 CPU, Jetson TX2 GPU and ZCU102 FPGA, using their vendor optimized vision libraries: OpenCV, VisionWorks and xfOpenCV. Our results show that the GPU achieves an energy/frame reduction ratio of 1.1-3.2× compared to the others for simple kernels. While for more complicated kernels and complete vision pipelines, the FPGA outperforms the others with energy/frame reduction ratios of 1.2-22.3×. It is also observed that the FPGA performs increasingly better as a vision application's pipeline complexity grows.

Towards Dynamic and Partial Reconfigurable Hardware Architectures for Cryptographic Algorithms on Embedded Devices

“…Power consumption is another major issue in the resourceconstrained embedded devices. It has been demonstrated [65], [66] that FPGA-based reconfigurable hardware often consumes less power than embedded microprocessor-based software-only designs. Furthermore, as stated in [60], [61], [62], [63], [64], the dynamic and partial reconfiguration could potentially lead to reduction in power consumption.…”

An Analysis of the Implementation of Edge Detection Operators in FPGA

“…It is worth noting that a comparison between FPGA and GPU implementations is beyond this work scope. Evidence from [11] and [23] proves that edge detection algorithms have a lower energy cost in FPGA, although GPU results are better when considering quality metrics.…”

“…Few works also evaluated the energy costs of edge detection filters, which is a crucial metric for embedded computer vision applications. For instance, the authors of [9] and [11] evaluated the energy consumption of hardware implementations of Canny filters, while the authors of [12] analyzed similar implementations of Sobel filter in FPGA. However, none of these works performed an analysis of different filters to compare their energy consumption.…”

Field Trial of a Flexible Real-Time Software-Defined GPU-Based Optical Receiver

“…With 45% [8] year-on-year growth of computing power and 25% increase [9] in energy efficiency (FLOPS per Watt), general-purpose GPUs have the potential to meet demanding processing requirements. Note that, GPU power efficiency showed a 3-fold improvement over equivalent FPGA for simple highly-parallelized operations [10]. These exponential increases may facilitate GPU use beyond prototyping.…”