A low-frequency array is being set up to observe the Deuterium hyperfine line at 327.4 MHz in Mauritius, in the southern hemisphere. The array will be used to measure the total power of the Deuterium emission in the local galaxy in order to determine the D/H abundance ratio. Radio astronomy is a compute intensive discipline. With the advent of large interferometers the digital processing is predominantly done using parallel computing hardware. So our project represents an opportunity to make use of off-the-shelf parallel hardware such as gaming GPUs to establish a low-cost high performance pipeline for a radio telescope. It also constitutes a test-bed for the development of interferometric imaging techniques on parallel computing platforms. For this purpose a parallel GPU FX correlator was designed using the cuBLAS library. We made use of a function performing the batched multiplication of matrices to compute the full correlation in the frequency domain. The cuBLAS implementation is compared with the xGPU implementation and Roofline graphs are plotted for both using different instrument characteristics. The cuBLAS-based correlator might not be faster but represents an optimised solution quickly implemented with a very short development time. The data rates and performance achieved will enable real-time processing for the telescope.

正在建立一个低频阵列，以观察南半球毛里求斯327.4 MHz的氘超细线。该阵列将用于测量本地星系中氘发射的总功率，以确定D / H丰度比。射电天文学是一门计算密集型学科。随着大型干涉仪的出现，数字处理主要使用并行计算硬件来完成。因此，我们的项目为利用现有并行硬件（例如游戏GPU）建立射电望远镜的低成本高性能管道提供了机会。它还构成了在并行计算平台上开发干涉成像技术的试验台。为此，使用cuBLAS库设计了并行GPU FX相关器。我们利用函数执行矩阵的批处理乘法，以计算频域中的完全相关性。将cuBLAS实现与xGPU实现进行比较，并使用不同的仪器特性绘制了Roofline图。基于cuBLAS的相关器可能不会更快，但是代表了一种在很短的开发时间内就可以快速实施的优化解决方案。实现的数据速率和性能将使望远镜能够进行实时处理。基于cuBLAS的相关器可能不会更快，但是代表了一种在很短的开发时间内就可以快速实施的优化解决方案。实现的数据速率和性能将使望远镜能够进行实时处理。基于cuBLAS的相关器可能不会更快，但是代表了一种在很短的开发时间内就可以快速实施的优化解决方案。实现的数据速率和性能将使望远镜能够进行实时处理。