We describe the implementation of a simple method for finding k approximate nearest neighbors (ANNs) on the GPU. While the performance of most ANN algorithms depends heavily on the distributions of the data and query points, our approach has a very regular data access pattern. It performs as well as state of the art methods on easy distributions with small values of k, and much more quickly on more difficult problem instances. Irrespective of the distribution and also roughly of the size of the set of input data points, we can find 50 ANNs for 1M queries at a rate of about 1200 queries/ms.