To be fair, there should not be a lot of data transfers to and from the GPU. Moving larger chunks of data (instead of many smaller ones) when you are running out of memory, using asynchronous data / compute streams would increase the performance.