Being an experimenter with my new Asus X99-E WS motherboard after running about 50 runs of PPBM8 and having four x16 slots which theoretically can run simultaneously at 16 lanes, today I could not resist running a third GTX graphics card. In slot 1, I installed my GTX 780, slot 3 had a GTX 770 and Slot 5 had my GTX 970. (slot 7 has an Areca ARC-1883i) and I have a Samsung XP941 which also uses 4 PCIe lanes.
Amazing all 3 Graphic cards contribute to the MPE acceleration! If my 860 watt power supply had more output connectors and could handle the load I would have tried a fourth card. My best score with our GPU assisted MPEG2-DVD timeline export was 13 seconds with two GPU's and with the third GPU added I got a 23% improvement to 10 seconds. Here is GPU-Z showing all GPU's are working simultaneously. I had to duplicate the sequence 4 times to give me time to collect data.
Premiere 126.96.36.199 and notice that the GTX 970 is working harder than the other two
I guess I should have said that the GTX 970 is trying to work harder than the other two. Notice that the Memory Clock on the GTX 970 SC which should be 1750 MHz according to the specifications falls back to 1500 MHz when running any CUDA application, while The GTX 780 SC is running at the its specified 1750 MHz and notice my GTX 770 SC is easily overclocked to the 2000 MHz
If you have not paid attention to these confusing GPU Memory Clock speeds before there is no consistency which speed is specified. For instance they might specify 1500 MHz one place (like GPU-Z) and 3000 MHz in another test like NV-Z and on some specification sheets say Effective Memory Clock 6000 MHz and/or 6 Gbps
MPE acceleration is dependent on CUDA cores and Memory Speed (Memory Bandwidth) and now on architecture. nVidia claims Maxwell is 1.4x more effective that Kepler was.
I am exploring the Maxwell memory speed reduction without success so far. But it is not an Adobe problem the same thing happens on any CUDA application and also is on the GTX 980.
I still believe the GTX 900's are a great performer for medium to high end cards.