Unigine Heaven was always very successful in finding issues for me.
From Kepler onwards, the card decides by itself what voltage it wants to use. It will go to its limit only if temperatures are fine and the power usage is below whatever is the target for that card. While you test, force the temperatures down by using an open case or turning up the case fans, and keep looking at the sensor view of GPU-Z or similar. If you don't do that, it won't use the highest possible clocks so you won't see if the chip is really stable or not.
What you could do after that is see at what point it starts to crash if you add a bunch of MHz to the clock offset. If there's headroom over there, you can be sure that it'll be stable at stock clocks.
I have a factory overclocked GTX 680 here with 4GB VRAM that was used for years of compute tasks, rendering stuff with Blender. It can't run stable at its default clocks after that torture, but it runs stable with a -40 MHz (minus forty) offset for the clocks, or if the BIOS is edited and set up more conservative like on the NVIDIA reference BIOS. Diagnosing the issues with that particular card, Heaven was the best and fastest program I could find for crashing it intentionally.
Originally Posted by Xinoxide
furmark can be dangerous though.
not only can it exceed the tdp and power draw of your card, this can exceed the limits of your PSU.
Use it at your own accord.
That's not possible for the cards the question is about. Kepler doesn't have a fixed voltage and speed, instead has a list of voltage+clock combinations. It will by itself select a less aggressive combination of voltage+clock to always stay below the card's limit for power draw. It also does that if the chip gets hot.Edited by deepor - 3/30/15 at 5:34am