I'm not really what he means either. If he was talking about kernel launches, there'd be no way he was getting 200k/sec since there's a bit of a launch latency (tens of milliseconds); the EngineYard program that one of the members posted on the CUDA forum can do over 200 million SHA hashes on a good card. One member said that he was getting over a billion hashes/sec on his multi-GPU setup.

By iteration, I mean one loop of my algorithm. I'm not brute forcing. There are a few other calculations I'm doing in addition to the hamming distance and the hash. I'm relying heavily on the random characters you get to append.