Thomas Unterthiner <thomas_unterthiner(a)web.de> writes:
Just FYI, this turned out to be a bug in cuBLAS that
only affects GK110
chips. According to nvidia this will be fixed with CUDA 6.5 (the bug is
still present in the current 6.5RC release).
Thanks for keeping the list updated while you were debugging this. It
*did* make me nervous, but it was great to hear that it was ultimately
not due to an issue with PyCUDA.