It has recently become practical to use triangular fault elements in earthquake simulators instead of the usual rectangular fault elements. A simulator spends most of its time computing how slip on some fault elements affects the stresses on other fault elements. We explore whether rectangles or triangles yield more accurate stress values on curved fault surfaces. One might expect triangles to perform better, because triangles can follow the shape of a curved surface, whereas rectangles leave gaps and overlaps between adjacent fault elements. Our test results show that, contrary to expectations, rectangles overall perform as well as or better than triangles when computing stresses on curved fault surfaces. We also find that one triangulation may perform significantly better than another triangulation.