Stochastic Gradient Descent with GPGPU

David Zastrau, Stefan Edelkamp

Abstract: We show how to optimize a Support Vector Machine and a predictor for Collaborative Filtering with Stochastic Gradient Descent on the GPU, achieving 1.66 to 6-times accelerations compared to a CPU-based implementation. The reference implementations are the Support Vector Machine by Bottou and the BRISMF predictor from the Netflix Prices winning team. Our main idea is to create a hash function of the input data and use it to execute threads in parallel that write on different elements of the parameter vector. We also compare the iterative optimization with a batch gradient descent and an alternating least squares optimization. The predictor is tested against over a hundred million data sets which demonstrates the increasing memory management capabilities of modern GPUs. We make use of matrix as well as float compression to alleviate the memory bottleneck.