Quote from: Morgahl on January 22, 2014, 09:27:30 AM

T kernel in Windows on my Titan is very unstable during autotune unfortunately anything that allocates more then 3GB of VRAM just crashes Cudaminer outright. Not sure what direct limitation is causing this but this is a consistent observation with several hours of manual configurations.

1. Any idea why there would be this limitation in the T kernel for Scrypt-Jane? If I set to autotune Cudaminer instantly takes up 5.5GB of VRAM but once it begins testing any config that would result in > 3GB to be allocated the autotune crashes out. Any manual settings to allocate more then 3GB falls under the same issue.

2. Why do you think the T kernel favors T64x1\1, T64x2\2, T64x3\3, T64x4\4 so heavily if it only allocates ~2.4GB of memory? I can take this as high as ~T75x#\# before reaching that 3GB allocation limit, but the hash rates are roughly half of the T64#\# configs? This seems to be counter intuitive to the "More memory allocation is better." thought process.

3. The K kernel similarly has a 3GB limit on the Titan but actually allows allocation in the ranges of ~K80x#\#, Commit 92 showed higher hasrates ~7% using K21x1 -C 1 over T16x1 which was consistently the best T kernel autotune?

4. Do you feel this 3GB limit may vanish using a Linux build due to better VRAM allocation?

I understand that you may not be able to address these issue directly via code changes (if it is determined to need a change here), mostly speculation and curiosity at this point. Thanks in advanced!