GPU runs of the code base have memory limitations. It would be interesting to compare profiling results from GPU runs to those from CPU runs, and see whether anything useful can be learnt from that comparison.