Massively Parallel EEG Algorithms for Pre-exascale Architectures

High-density EEG is a non-invasive measurement method with millisecond temporal resolution that allows us to monitor how the human brain operates under different conditions. The large amount of data combined with complex algorithms results in unmanageable execution times. Large-scale GPU parallelism provides the means to drastically reduce the execution time of EEG analysis and bring the execution of large cohort studies (over thousand subjects) within reach. This paper describes our effort to implement various EEG algorithms for multi-GPU pre-exascale supercomputers. Several challenges arise during this work, such as the high cost of data movement and synchronisation compared to computation. A performance-oriented end-to-end design approach is chosen to develop highly-scalable, GPU-only implementations of full processing pipelines and modules. Work related to the parallel design of the family of Empirical Mode Decomposition algorithms is described in detail with preliminary performance results of single-GPU implementations. The research will continue with multi-GPU algorithm design and implementation aiming to achieve scalability up to thousands of GPU cards.