ION GNSS 2012
Session E4: Software Receivers

Title: Performance Analysis of GPU Based GNSS Signal Processing
Author(s): U. Haak, H-G. Busing, P. Hecker, TU Braunschweig, Germany
Date/Time: Thursday, September 20, 2012, 4:23 p.m.
Room: Grand Ballroom West (Renaissance)

The increasing number of developments of software-based GNSS receivers especially in the research area has been initiated by two major aspects: The first aspect is the lack of interfaces in off the shelf receivers which are necessary to realize positioning solutions based on a deep sensor integration. The second aspect is the increase of computational speed due to the developments in the computer sector in the past years which made real-time software receiver applications without additional processing hardware like FPGS boards possible.

The Institute of Flight Guidance (IFF) developed a software receiver that uses graphic processing units (GPUs) to achieve real-time capability. The software receiver is programmed in C++ and is currently running on windows maschines only. It is able to track GPS L1 and SBAS L1 signals and provides measurement data with up to 50 Hz in RINEX format and positioning solutions via ethernet connection to a client GUI program. The receiver is specialized on nVIDIA GPUs using the proprietary Compute Unified Device Architecture (CUDA) by nVIDIA. Using their massive parallel compututational speed, real-time capability is possible depending on several factors like the sampling rate of the front-end, the number of tracking channels or the selected GPU board. The receiver is designed to be flexible with respect to the front-end. In the current implementation three different kinds of receiver front-ends with different quality classes can be used. Furthermore, different graphic boards with significantly varying processing performance are used, starting with the Quadro FX 370M, which is designed for notebooks, with only 8 CUDA cores and ending with the GeForce GTX 480 with 480 CUDA cores.

Generally one crucial question is the performance of the signal processing implementation. For several applications it might be of interest to minimize the computational requirements. The IFF is currently involved in research projects in the automotive sector with focus on sensor integration. For development and testing of algorithms for automotive applications the IFF operates the software receiver in a research vehicle. A Volkswagen Passat station wagon has been customized for this purpose. It is equipped with an inertial measurement unit as a reference system. Furthermore the Controller Area Network (CAN) of the power train and the assistance systems are mirrored and can be accessed in real-time. The vehicle is equipped with an automotive PC for data processing. The choice of the graphics board for signal processing is limited by the available system power that is significantly lower than the available power in standard desktop computers. The core design goal is to optimize the signal processing source code to enable real-time speed in the automotive system environment.

The latest CUDA release offers a profiling toolkit to assess the performance of processing implementations. The tracing funcionality of the toolkit allows to review the processing chain and may expose unnecessary idle times in data processing. The device memory useage and data transfer bandwidth can also be evaluated using the profiler. Using these tools the software receiver has been optimized for deployment in the vehicle PC.

In the proposed paper the influence of different hardware and software configurations as well as optimizing possibilities will be presented. Furthermore the detailed profiling results will be discussed especially regarding the workload ratio of the individual receiver threads with respect to the data handling overhead.

Previous Abstract Return to Session E4 Next Abstract