CUDAICA: GPU Optimization of Infomax-ICA EEG Analysis
Figure 2
Execution time of Infomax step using ATLAS, MKL, and CUDAICA for 30-minute experiments (a). This short experiments may last more than 4 hours of computation time in the standard BLAS implementation with 256 channels. MKL shows significant performance improvement over BLAS, and CUDAICA behaves the best. Comparison between ATLAS and MKL versus CUDAICA (b). CUDAICA shows a maximum speedup of 4.7 over MKL, and more than 20x over standard ATLAS implementation. In (c), we show the time per step of different implementations of Infomax with several channel configurations, running under the available hardware.
