It's a GTX 670. has 1344 cuda cores. 1344x32 works at -L 512. No matter what I specify for -L, doubling doesn't work. Out of memory every time.
but can't you just ignore the out of memory? it's for scrypt related buffers mostly, which keccak doesn't need.
maybe -L 1024 -m 1 does the trick?
It appears that this still works the best for me on a GTX 780
cudaminer --algo=keccak -o stratum+tcp://maxpool.1gh.com:17333 -u ma4uR1HJpGZy3jZwTWn2Q5ivA7zS31ZuEC -p x -m 1 -d 3 -L 1024 -l K2304x32 -i 0