Following the computation on total memory used, what is the reason that not even half of the memory available on the Titan is being used? Up until the last release today (4/14) autotune has always picked a configuration of 300-307x2 (now 263x2) which comes out to about 2.5GB of used memory at most. Since scrypt is a space/computation tradeoff algorithm isn't there potential for almost a doubling in hash performance if the full 6GB could be used?
At the moment there is no trade-off. I use the full scratchpad size per thread. So required memory size scales with the number of threads.
A trade-off happens when you try to reduce the scratchpad size at the cost of increased computation. Trying this is on my TODO list.
Christian
Ah, so if you're using the full scratchpad space right now do you have an idea why I can't use any launch configurations that go over 2.8GB? The program says there isn't enough memory, but according to the calculation there should be. I'm wondering if there is an integer overflow comparison somewhere?