Nvidia 580+770 strange render results

I have recently added an nvidia gtx 770 card to my machine, joining my gtx 580. I have been testing the cards and am trying to understand the results.

My test scene is 500 samples, which the 580 renders in 4:18. The 770 renders it in 4:22. I am using tile size 512x512. I have tried other sizes for the 770, including 256x256, 512x256, 128x128, and many more, but 512x512 does the best. I do not understand why this is the case. The 770 has triple the CUDA cores of the 580. Why isn’t there more of a speed improvement? Is there some magical tile size I’m missing?

I also ran both cards working together. The option in preferences says 580+770. Together, they render the test scene in 2:58, so that is a nice speed boost. However, I have been running a GPU monitor (gpu shark) and I noticed strange numbers. Most of the time, the 580 is at 99% GPU used, while the 770 sits at a mere 1-10%. Occasionally the 770 goes up, but only when the 580 is going down. It’s not 50% and 50% either, more like 30/30 or 60/20, always in favor of the 580. If I turn down the tile size to 256x256 or 128x128, then I see the 770 with a higher GPU % use in comparison, but render times actually get worse by a few seconds.

I am concerned too because GPU shark is almost always showing Voltage limit reached under Limiting Policies for the 770. The 580 has no power information displayed at all, and never shows any limiting policies. I haven’t been able to figure out what this actually means, if anything. I have a 1200w power supply so I know I’ve got plenty of juice. Is this limiting policy throttling the card or something?

Can anyone help me make sense of this? I got the 770 for the extra VRAM, but I was expecting more of a boost in render speed than this based on benchmarks I’ve seen.

The GTX 770 is basically a rebranded and slightly overclocked GTX 680. From the GTX 5xx series to the GTX 6xx series we saw a dramatic decrease in performance per CUDA core, which made the GTX 680 in fact slower than the GTX 580 was. So, in a nutshell: You can’t compare CUDA cores between different series and the results you see were to be expected.

Well, that’s a shame, but not a big deal. Its not like there are other 4gig card options short of the 780 or 980. Thanks for the response.