DescriptionWe aim to mitigate the performance bottleneck of migrating data between host and device memory in GPU applications by accurately predicting application access patterns using deep neural networks. We model the memory access pattern of any given application by collecting page faults that trigger data migration to the GPU and feed this time series as input to a neural network that outputs the next several page faults. We evaluate these predictions on the basis of what makes a useful prefetch in our context for the GPU. Current work has looked at trivial GPU applications, such as matrix operations, and moving toward real, complex applications. Our work will be presented by bringing attention to the predictive capability of our neural network on the current applications tested.