If you don't feed the CUDA miner specific gpugrid and gputhreads parameters, it will do a benchmark of all of the available options to find the best one.  Either due to a bug or driver issues, some of your first couple of runs of the CUDA miner may be less than you could get.  Sometimes if you leave your system idle for about 60 seconds, then you get better speeds.  Sometimes, you would have to run it immediately after closing it to see the best speeds.  It's somewhat random.  If you do see one instance of it running the fastest, take note of the last pairs of numbers that show up.  For me, the best pair of numbers ended up being (128,128).  They'll show up in the order (gpugrid, gputhreads).  Otherwise, you could just run the miner yourself, constantly varying the inputs until you see the greatest average mining speed.  The first report of the hash speed will be lower than all of the others, so give it some time.  Also, for the large majority of nVidia cards, your ask rate should be larger than the default.  If you get less than 40 Mhash/s, this should be 10 seconds (10000 ms) or greater.  Everybody has a different theology for the interval between getwork requests, so if you mine in a pool, you may have to PM a pool operator for the settings that they would like you to use.  They'll need to know your hash rate.  As far as system slowness, there's an aggression setting you can raise or lower to change the responsiveness.


I haven't had the time to run it, yet.  I'll be doing that sometime within the next 24 hours.  I'm glad to test it out for you and offer any suggestions I can, since I've been toying with these miners for some time.


Yeah, I can't see how you would be able to fix this other than making a specific form for each miner.  As far as I know, puddinpop's RPC OpenCL miner and m0mchil's Python OpenCL miner yield pretty close to the same hash rate on a majority of hardware.  I don't think there will be too much of an outrage if puddinpop's OpenCL miner gets left out, but it would still be a little bit of a headache to set up hardware lists for each miner (OpenCL vs. CUDA).  The RPC CPU miner that does not have the -4way suffix runs the same code as the bitcoin client, if I remember correctly.  That one can probably be ignored, as well... maybe.  Also, CUDA has this weird behavior of making my secondary card be GPU 0 and my primary be GPU 1.  Just thought I would pass that along.