A bitwise reproducible deep learning framework

Is there any indication of what the performance hit for this might be?