AVX faster than AVX2 for some gravity interactions.
See the closed MR !1763 (closed)
Would be good to see if autovectorization differences between the architectures could be improved and the performance at low-z repeated with AVX2.
Edited by Peter W. Draper