Nvidia 350.xx with opencl 1.2

[n_w95482]

I tried 350.05 a few days ago with my GTX 680. Same huge drop in performance (using FAHBench to test) as any other driver above 327.23. Sigh...

[folding_hoomer]

If there is only OpenCL 1.2 - support for Kepler and Maxwell, there is no benefit for the older GPU´s (especially Fermi).
So - if used by PG, all Fermi-based GPU´s were "out of the game"

[kscott]

If there is only OpenCL 1.2 - support for Kepler and Maxwell, there is no benefit for the older GPU´s (especially Fermi).
So - if used by PG, all Fermi-based GPU´s were "out of the game"

Your answer makes no sense
The 680 is Kepler, not Fermi

[bruce]

You're spending a lot of time worrying about a problem we don't have.

As I said at the bottom of the previous page, FAH will be slow to adopt OCL 1.2 because they'd have to find a way to continue to support those Donors who only have access to an older version. It's possible to create a client that recognizes which version you have and to create FAHCores that are not compatible with older versions and add enough logic to the WU distribution system to restrict projects from sending out WUs that you can't process, but ... it's in the "not soon" category because that's a lot of changes that need to be coordinated.

[7im]

It's also possible to have multiple code paths so that a core uses the older version on older cards (already exists) and uses the new code path on newer cards, if OpenCL 1.2 is present, and if it brings any functionality or speed to fah. Like a CPU core using SSE, or SSE2, SSE3, which ever is faster and supported, etc.