Thanks! Yes this is exactly it. Interesting about the choice of 32 bits for carries. N=67 is pretty high. I suspect N could be tuned down significantly depending on the problem. Eg global sums in geophysics models -- the range of values being summed does not span the full double range. But that would require analysis to decide about, while the full sized superaccumulator works directly. I wonder how it stacks up in terms of performance w/r/t the methods in the blog post