Neuroimaging Registration on GPU: Energy-Aware Acceleration.

We present a CUDA implementation for Kepler and Maxwell GPU generations of neuroimaging registration based on the NiftyReg open-source library [1]. A wide number of strategies are deployed to accelerate the code, providing insightful guidelines to exploit the massive parallelism and memory hierarchy within emerging GPUs. Our efforts are analyzed from different perspectives: Acceleration, numerical accuracy, power consumption and energy efficiency, to identify potential scenarios where performance per watt can be optimal in large-scale biomedical applications. Experimental results suggest that parallelism and arithmetic intensity represent the most rewarding ways on the road to high performance bioinformatics when power is a major concern.