A Tradeoff Analysis of FPGAs, GPUs, and Multicores for Sliding-Window Applications

Abstract

The increasing usage of hardware accelerators such as Field-Programmable Gate Arrays (FPGAs) and Graphics Processing Units (GPUs) has significantly increased application design complexity. Such complexity results from a larger design space created by numerous combinations of accelerators, algorithms, and hw/sw partitions. Exploration of this increased design space is critical due to widely varying performance and energy consumption for each accelerator when used for different application domains and different use cases. To address this problem, numerous studies have evaluated specific applications across different architectures.In this article, we analyze an important domain of applications, referred to as sliding-window applications, implemented on FPGAs, GPUs, and multicore CPUs. For each device, we present optimization strategies and analyze use cases where each device is most effective. The results show that, for large input sizes, FPGAs can achieve speedups of up to 5.6× and 58× compared to GPUs and multicore CPUs, respectively, while also using up to an order of magnitude less energy. For small input sizes and applications with frequency-domain algorithms, GPUs generally provide the best performance and energy. . 2015. A tradeoff analysis of FPGAs, GPUs, and multicores for sliding-window applications.

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

“…Other comparison studies focused on a subset of vision kernels. For example, the study in [7] and [8] evaluated the performance of sliding window applications on FPGAs, GPUs and multi-core CPUs. They compared the performance of three applications: Sum of Absolute Differences (SAD), 2D convolution, and correntropy.…”

Benchmarking vision kernels and neural network inference accelerators on embedded platforms

“…Other comparison studies each focused on a subset of vision kernels. For example, the study in [9] and [10] evaluated the performance of sliding window applications on FPGAs, GPUs and multi-core CPUs. They compared the performance of three applications: Sum of Absolute Differences (SAD), 2D convolution, and correntropy.…”

Boosting the Hardware-Efficiency of Cascade Support Vector Machines for Embedded Classification Applications

“…A hybrid FPGA-GPU pedestrian detection is presented in [40] where the SVM is implemented on the GPU and a feature extraction algorithm on the FPGA for 800×600 images and achieves over 10 frames-per-second for the classification of 1000 windows. However, GPUs are power hungry devices compared to FPGAs [29], [41], (FPGAs consume approximately an order of magnitude less power as shown in [13]) and as such they are not suitable for power-constrained embedded applications. In addition, existing GPU implementations do not translate well to the more energy-efficient embedded GPUs due to less available resources [42].…”

“…Relevant research has produced quite promising results in terms of performance for use in embedded systems. However, even if implementations of SVMs on GPU platforms have gained considerable attention due to the high-level programming capabilities, GPUs still face challenges with regards to power consumption, especially with the increasing development of portable resource-limited platforms, requiring specific hardware solutions and large scale problems [12], [13]. Hence at present, computing systems based on FPGAs and customized hardware accelerators allow to exploit the inherit parallelism of algorithms such as SVMs, whilst achieving efficient implementation suitable for real-time processing and low-power operation.…”