So not quite analogous to PIN, but Nvidia does offer a profiler to run at the same time as your CUDA programs to give you perspective on how well the GPU is being utilized. It looks like OS X's Activity Monitor on steroids. The Analysis and Timeline views look like they could be enormously useful in the future (and would have been a huge help on Assignment 2.) nvprof is runnable from ghc machines, although I couldn't get the program to launch over SSH. If anyone has success with it, do share! See here the console output of running nvprof ./render -r cuda rgb on the gates machines:

==4775== Profiling application: ./render -r cuda rgb
==4775== Profiling result:
Time(%)      Time     Calls       Avg       Min       Max  Name
 88.50%  14.425ms         8  1.8031ms  1.6160ms  2.7931ms  [CUDA memcpy DtoH]
  9.17%  1.4951ms         7  213.58us  213.31us  214.05us  kernelRenderPixels(void)
  2.27%  370.31us         7  52.900us  52.704us  53.376us  kernelClearImage(float, float, float, float)
  0.06%  9.3440us         9  1.0380us  1.0240us  1.1520us  [CUDA memcpy HtoD]

This comment was marked helpful 1 times.

And in case anyone was wondering, to get the visual version of the profiler as pictured, run nvvp over an X session.

This comment was marked helpful 1 times.