It appears similar to the disaster of ntlm, that the md5 OpenCL may have the same issues. I *thought* we had run this algorithm before so it may not be the algorithm but issues with the hybrid2 code. In any case I have halted the production of new work for the GPU side, though there is still work already in the buffer.I think in the immediate case we have the following to do ASAP:1) find a way to continue Nvidia CUDA even if we have to disable OpenCL AMD/ATI (or vice/versa) so it isn't a 100% GPU shutdown2) fix the OpenCL codeI am about to go pick up a new GPU that will let me directly attach vms in my server to the card, such as windows x64. On the GPU side PowerBlade has had the burden of all things windows related, and it is past time that I have a dev environment for both sides. I do have notes in the code for the ntlm side that has a very similar design on the md5 side that might be the issue in both, but I will know more as soon as I can test that directly.*If* we could come up with a table construction for mysqlsha1 to keep work flowing in the meantime that would be great for crunchers, but we have exhausted the algorithm to a large extent with the code currently deployed. In the mean time work is likely to:1) dry up2) erroneously validate in some cases 2 OpenCL results agree (we still do limited checks on the server even with matching WUs)3) result in a lot of inconclusive resultsIn any case further updates will be made in as timely a manner as possible. I do not see the time-frame to fix the OpenCL code bugs and get that code out to clients to be more than days.

More...