Implementation of large-scale fir adaptive filters on NVIDIA GeForce graphics processing unit

This paper presents implementations of an FIR adaptive filter with a large number of taps on nVIDIA GeForce graphics processing unit (GPU) and CUDA software development environment. In order to overcome a long access latency for slow off-chip memory access, reduction of memory accesses by re-ordering and vector load/store operations and an increase of the number of threads are introduced. A tree adder is introduced to reduce the cost for summing thread outputs up. A simultaneous execution of multiple filters are also examined. On low-cost platform such as an Atom/ION nettop, GPU will accelerates the computation by almost three times. For simultaneous multiple simulations such as an ensemble averaging, a GPU with a large number of processing elements outperforms a dual-core CPU; almost six times faster for 16 runs.