I get a CUDA-based program, and it can run on Geforce GT 650M and windows 8 with CUDA 5.5. The same program with CUDA 5.5 also can run on Kepler 20m and windows server 2008 R2. When I increase the amount of computation, the computer with kepler 20m and server 2008 R2 restarts before the computation finishes. However, the same computation can run correctly on Geforce GT 650M and windows 8. Thus, I think the program is correct.I have no idea about why this happens.
Sounds like either your power supply is delivering insufficient power and/or the PCI-E cables that power the card are defective. K20m can use up to 225W as opposed to your GT 650m which doesn’t come close to that figure, so it makes sense that you get a restart if you are having power issues at high watt loads.
I use the nvidia-smi.exe to monitor the GPU state. I find that the temperature of the k20m increase to 95 degree Celsius, and the computer restarts. I think this may be the reason.
High temperatures would do it as well, yes. Better cooling would be the solution.