Low arithmetic intensity means that there just aren’t enough arithmetic operations requested for each data item (maybe one addition and one multiplication), consequently performance is memory bound. That also means that you can be doing additional arithmetic for free. Useful arithmetic includes things such as
Kahan compensated summation or blocked summation which can radically reduce round-off error of large summations. A recent paper showing the improvements possible is http://eprints.maths.manchester.ac.uk/2704/1/paper.pdf
1 Like