kernel_security_check_failure

I’m running two Tesla K40c cards on a Windows 10 machine with the latest driver upgrade for cuda 7.5 toolkit. My neural network software works fine until the matrix gets to a certain size. 101 X 166, using unified memory. I’m only using one card as of now, and the utilization of that card is around 85% when I get a BSOD stating KERNEL_SECURITY_CHECK_FAILURE.

The dump says "Probably caused by nvlddmkm.sys (nvlddmkm+3c232)

Any ideas on how to fix this problem?

array of size 101 X 166 works, but 101 X 167 crashes.