This
document provides some insight on optimal OpenCL performance of nVidia cards.
One thing that MANY users seem to not grasp is the total memory management of a system as they work together.
In Terms of OpenCL, you can have a GeForce Titan on any given system and perform no better than having a GTX280. Since the the GTX 280 has a 512 bit memory interface at half the bandwidth (which is at 141gb/sec, not bad considering the GTX650Ti nearly the same at half the interface width.
Now those numbers are theoretical and never reach near those numbers. Mainly because if you're not running a socket 2011 processor your cpu bandwidth is only 25-29 Gb/s. On top of that, in terms of OpenCL bandwidth usage, the cpu can only achieve that bandwidth as long as ALL the cpu's memory channels are saturated.
For example; If you have a 6 Core cpu with multi-threading (12 total). 8GB or even 16Gb of RAM is not nearly enough memory to maximize the cpu's throughput. At 8Gb the CPU is throttled with less than a GB of RAM per core. Each core needs at least 2gb per, 4gb for best performance in terms of OpenCL. You'd need 24Gb of RAM minimum, 48GB to manage the bandwidth of a Titan. Again, this is in terms of OpenCL performance.
In contrast, with all RAM and CPU being equal, the AMD HD7970 has a memory bandwidth of ~240GB/s at a 384 bit wide interface. That's 58% more bandwidth than the GTX650 Ti or the GTX280. That's a tremendous difference even if the GTX280's interface is 512 they even each other out because of the GTX's higher bandwidth. The Titans interface is the same as the AMD 7970's at 384 bits. The AMD 7970 has 25% lower bandwidth than the Titan and should in all cases be slightly slower than a Titan.
I know this sounds like gobbledygook. But it makes sense.
Studio1productions.com goes into great detail about this very subject. I chose to compare Titan specs because it's the top non-Quadro card from nvidia. Seeing that Quadro drivers aren't specific to OpenCL. OpenCL uses ALL resources no matter what drivers are being used.
