The reason there is no runtime improvement due to changing Par0 from 2 to 4 is because the our computation becomes bandwidth bound.

My understanding of "Change increases resource utilization without improving performance" is that we can do more work faster by making the outer loop more parallel, but this doesn't speed up the overall program because DRAM transfers is now the limiting factor.

@wooloo great summary. That means even if the outer loop finishes quicker, it needs to wait for the DRAM requests.