Do CUDA cores stack in SLI?

ShindoSensei

Honorable
Mar 6, 2013
147
0
10,690
I am planning to buy parts for a future rig, and will buy the 700 series, but I'm picking out reference parts for now. In SLI do the CUDA cores stack? I know it may not be exactly double as some of the performance could be lost, but I'm considering buying two GTX 680 4GB editions (for reference), and will water cool them, as EVGA's hydro copper edition only has 2GB of VRAM. If the cuda cores do stack I will buy 680s, if not I'll pick up a TITAN because apparently the GTX 690 can only use 2GB of its vram because it doesn't stack in SLI. Thank you!
 
when it comes to SLI, Vram does not stack, and work is shared between the two cards. You cannot SLI cards that are not the same, as SLI tends to be slightly more strict when it comes to its requirements. CUDA cores and Vram are 2 different things.
 
Currently programs have to be written to take advantage of multiple CUDA devices, SLI doesn't expose more CUDA cores as a single logical device. Same goes for OpenCL.

While you're likely to see this kind of behavior from things like bitcoin mining software, you're unlikely to see it in general consumer programs.
 


I know, which is why I believe the gtx 690 does not really have 4gb of usable vram. However my question is if CUDA stacks or not.
 


So would my best bet be to get the TITAN hydro copper? It's clocked higher than the 690, and has 6 gigs of vram, which should help for newer games with tons of stuff.
 
If you can pick a single GPU solution that performs on par with a dual GPU for the same price, go with the single GPU. Titan will give more consistent performance compared to GTX 690.
 
This is a topic of some interest to me. I want to video edit using PP6/7 and with the Mercury engine it's all about CUDA. I have my eye on a matched pair of GTX 680. IF I SLI them will I effectively have 3072 cores in play?
Appreciate any assistance as this is my first ever build (I'm 62 and I don't want to mess it up)
 

Latest posts