Mail from Thomas today:

“…with latest SVN this surprisingly works?! Simply enabling the feature and compile the CUDA kernel (sm_21 tested here, on Geforce 540M). As far as I can tell it renders correct with all fallof modes. Have not messured performance yet.”

Confirmed by benchmark tests on BA (win64 test build available) – definitely faster than CPU and seems to have no issues. Magic! It works and we have no idea why! Perhaps Brecht can spread some light on it this Sunday, for now I shall believe the path-tracing gods are at work.

However we can be pretty sure it won’t make it to 2.69, since that’s currently in feature-freeze, bug fixes only. But 2.70 is a good bet!