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 ou

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.