Dont want to reinvent the wheel testing, if the answer is already out there somewhere .... definitive not anecdotal.
Is PPS Sieve GPU app sensitive to GPU memory bandwidth?
For example, with Collatz (needs memory high) the datasets going back and forth CPU-GPU are large and need all the GPU memory bandwidth you can give it, there is a very noticeable difference in output if GPU memory is set at (say) 500 compared to set at (say) 1250 or more.
With MilkyWay(ideally memory low), its the opposite. The datasets are small, and GPU memory bandwidth setting is irrelevant - it can go as low as 180 with no discernable difference in output. Running MW at (say) 1150 or more, is a waste of power, and only produces a more efficient space heater.
So for PPS GPU Sieve - memory high or low?