MSc Dissertation: Andrew Woods


Adobe-PDF-downloadWoods, Andrew. Accelerating Software Radio Astronomy FX Correlation with GPU and FPGA Co-processors. MSc Dissertation. Department of Electrical Engineering, University of Cape Town, 2010.



This thesis presents two implementations of a frequency domain radio astronomy correlator. One implementation uses reconfigurable hardware (Xilinx Virtex 4LX100) and the other uses a graphics processor (Nvidia 9800GT). The objective of a radio astronomy correlator is to compute the complex valued correlation products for each baseline which can be used to reconstruct the sky’s radio brightness distribution. Radio astronomy correlators have huge computation demands and this dissertation focuses on the computational aspects of correlation, concentrating on the X-engine stage of the correlator.

Although correlation is an extremely compute intensive process, it does not necessarily require custom hardware. This is especially true for older correlators or VLBI experiments, where the processing and I/O requirements can be satisfied by commodity processors in software. Discrete software co-processors like GPUs and FPGAs are an attractive option to accelerate software correlation, potentially offering better FLOPS/watt and FLOPS/$ performance.

In this dissertation we describe the acceleration of the X-engine stage of a correlator on a CUDA GPU and an FPGA. We compare the co-processors’ performance with a CPU software correlator implementation in a range of different benchmarks. Speedups of 7x and 12.5x were achieved on the FPGA and GPU correlator implementations respectively.

Although both implementations achieved speedups and better power utilisation than the CPU implementation, the GPU implementation produced better performance in a shorter development time than the FPGA. The FPGA implementation was hampered by the development tools and the slow PCI-X bus, which is used to communicate with the host. Additionally, the Virtex 4 LX100 FPGA was released two years before the Nvidia G80 GPU and so is more behind the current technologies. However, the FPGA does have an advantage in terms of power efficiency, but power consumption is only a concern for large compute clusters. We found that using GPUs was the better option to accelerate small-scale software X-engine correlation than the Virtex 4 FPGA.




Leave a Reply