EDIT: I was trying to use the tesla kernel. It doesn't appear to like Jane. I can squeeze out 2.3 KH/s on my gtx 780 using K9x2. Going to keep tinkering.
2.4 kHash with K4x4 on GTX 660Ti, I also use -C 1
Tesla, Fermi and Legacy Kernels don't do scrypt-jane yet.
"GPU #0: Given launch config 'K4x4' exceeds limits for 1D cache." No warnings about -C 2, but the results don't validate.
6x3 and 9x2 with -C 0 give me the best results at ~2.3 and ~2.35 kh/s. Unfortunately my machine becomes annoyingly slow with any scrypt-jane config as it stands. With K9x2 I have 40MB of free gpu memory, with K1x1 over 2GB are free, yet my computer still approaches unusable, and as bathrobehero said, -i 1 doesn't seem to help either.
I'm thinking of picking up a 4GB GT 630 for $30 to play around with. It seems to match your criteria for being efficient. 96 shaders, 128 bit bus, lots of ram. I'm curious what kind of hash rate it can pull, but mostly I want to play around with scrypt-jane without crippling my dev box.
On another note (regarding regular scrypt), I've just been using the 12-18-2013 commit from github until yesterday... but with the changes from the 20th (autotune up to 32 warps) I get another 20 kh/s with T15x32. I played around with the Kepler kernel and noticed -C 1 adds nearly 50 kh/s. Is the texture cache a possibility for Tesla kernels in the future?
cbuchner1, you are a beast!
I concur. I will have to find some LTC to send your way to show my appreciation.