Hi Raphael. It's an interesting approach, and thanks for making it available. I just tried running your benchmark on an Intel Sandy Bridge system (E5-1620) and got results similarly disappointing results to those that you found on Haswell. I posted the numbers as an issue on Github.