http://www.cryptohaze.com/forum/viewtop ... p=264#p264
It's much, much faster on large hash lists.
On 5M hashes, 216SP GTX260:
Code: Select all
Work: 222 Done: 0.07% Time: 481.00 ms Stp/srch: 60.5M/s 302471968.0M/s
The Windows code has a few issues - the "FAST[hash]" stuff, which is a port of the 0.61 code, doesn't work. As a result, you'll want to specify longer kernel execution times (-m 100 or -m 500), and the auto-tune doesn't work (so manually specify larger thread/block counts, maybe --threads 256 --blocks 128, play around & see what's fastest).
I'm working to resolve the remaining issues, but wanted to get this out here and let people beat on it a bit. It will also be going open source shortly, as I get (yet another) complete code rework done to make it object oriented (not the kernels, but the host code), and build a TCP/IP daemon for it that will allow for multi-GPU support, distributed cracking (on a local scale, though there's no reason it wouldn't work on the internet with enough bandwidth), and will allow people to write other modules for it - as the protocol will be documented, if someone wanted to write an ATI module for it, it would work just fine (I may talk to mrb about this).