Exploring FPGA-GPU heterogeneous architecture for ADAS: Towards performance and energy

Abstract

This paper investigates the feasibility of using heterogeneous computing for future advanced driver assistance systems (ADAS) applications. In particular, we take lane detection algorithm (LDA) as a test case. The algorithm is customized into FPGA-GPU heterogeneous implementations which can be executed in either workload constant or balanced scheme. Then the heterogeneous executions are evaluated in view of performance and energy consumption, and further compared with the single-accelerator run. Experiments show that the heterogeneous execution alleviates both the performance and energy bottlenecks caused when only using a single accelerator. Moreover, compared with the single FPGA execution, the workload balance scheme increases the performance by 236.9% and 42.9% on our two tested platforms respectively, while ensuring the low energy cost.