Thanks everyone for the feedback and useful information.
I’m sorry if I understood wrongly some stuff, like the fact that cycles is getting slower over time, (that is, if we ignore that blenderguru chart). And I’ll remove the “ultimate” from the thread title to reduce some misinterpretations from other incautious users.
This makes me think, if I have a scene then it appears would be better to render the diffuse/glossy,etc on a pre-2.7x version of blender and separate passes for sss and volumetrics if necessary only on 2.78+ then composite it all, being a faster workflow than straight up rendering it on 2.78+.
A few years back we made a test with several Blender versions, just to see if they are getting faster or slower. At that time, the conclusion was that the newer versions were getting faster. We had the advantage of being able to use the same testing method and the same hardware, which made the results consistent. You can see the results here, if you’re interested.
Looking back at all the testing we’ve done over the years, there are a number of variables that are influencing the results: Blender version, environment (OS version, drivers, CUDA version, even the storage type - ssd or hdd), method of testing, etc. Because of that, all the benchmarks submitted by users will have some degree of variation in the numbers. Averaging them over time can help form a general idea about each card’s performance, but keeping in mind those variations will help put the numbers in perspective.
That being said, the 580 does hold its own. Too bad about the memory, though…
Marius