I have three GTX 670s in SLI. A few weeks ago, one GPU was failing with graphical glitches the moment I got into windows so I RMA'd it and got the replacement today. In the meantime, I was running 2-Way SLI just fine. Today I put in the replacement, installed drivers, enabled SLI, etc. then went to play Sleeping Dogs. During gameplay the framerate became unstable and the computer froze up and I had to do a hard reset. I was using the latest beta drivers so I assumed it was a problem with them. I went back to using the WHQL drivers that were installed before when I was using 2-way SLI while the third card was being RMA'd (314.22). So I tried to play Sleeping Dogs again and the framerate seemed much smoother, but then when I got to about the third fight segment the game suffered a massive slowdown. The framerate dropped to 0 then hit 200-something and continued jumping around as the game lagged to a near standstill and I was forced to hard reset again. At this point I figured I should run a benchmark or something to see if it happens again. So I started up Unigine Valley and began a benchmark at the highest settings possible at 2560x1600. Sure enough, after some time had passed it began again. This time I noticed that somehow SLI had disabled itself and only GPU 2 (the replacement card I just installed) was active. The framerate was jumping again and it was again lagging to a near standstill.
 
I enabled hardware logging during this. Here's the log file. You can see at [21-05-2013 21:45:03] where I had Valley running and the framerate hit 0 then started jumping as well as the usage for GPU 1 and GPU 3 hitting 0 with only GPU 2 (the RMA replacement) being active.
 
So I need to RMA this again? What on earth causes this anyways? I thought these replacements were tested.
Going to try some more tests, but I think this replacement card is broken.
 
EDIT:
I just installed a different SLI bridge I had and ran through the entire Valley benchmark twice without issue. I guess it was the SLI bridge which makes sense since SLI was randomly being disabled. Hopefully I don't encounter the issue again, if I do I'll no doubt be back with another hardware log.