Folding Forum

Anyway to tell when 8018 will be wrapping up? My GPUs are on a death march...

Something I have noticed....

I have a 560ti 448 and a 650ti folding next to each other in the same machine.

On core15 projects the 560 is about 30-50% faster than the 650. On core17 projects the 560 is only about 10-15% faster.

Is Kepler not well supported in core15?

Both are working through 8018s now...

560ti 448 TPF 5:20
650ti TPF 7:09

The short answer: Yes.

When Core15 and the appropriate WU´s were programmed there were no Kepler GPU to test and optimize the code for them - so it was only optimized for the than new GPU´s called Fermi . . .

That's not quite the answer. 1 Fermi core equals 2 Kepler cores in regards to CUDA programming (core 15) and the architectural differences of F vs. K. OpenCL in core 17 is more 1 to 1.5 in Fermi vs. Kepler.

If this were simply a matter of having newer cards around to use for optimization, then core 17 on Kepler should kick all Fermis to the curb, because core 17 was developed with newer cards available for testing. And since the 560 above still beats the 650, that pretty much proves optimizations are not the issue.

More importantly, if more speed was available from a core_15 optimization, Pande Group would jump on it, as they always do. They want more speed as much as anyone.

What is not clear in this discussion is how much of the "optimization" you're talking about is part of the FahCore and how much is in the drivers.

At one time, somebody wrote drivers that worked efficiently for the G80/Tesla hardware. The hardware guys made some improvements, called them Fermi, and sent them back to the driver guys and told them to update the drivers so they would work with either G80/Tesla or with Fermi along with a mandate that Fermi had to be more efficient and no performance changes could be made to G80. That's not possible, so they released drivers that were only slightly less efficient for G80. The same thing happened again when Kepler was introduced. The Fermi drivers necessarily took a (small) performance hit to provide a new path through the driver code that supports Kepler. We don't have enough information to figure out if the driver change was more or less significant than the FahCore optimizations because all we can measure is the sum of two terms.