flop counts of all operations to better be able to compare algorithms
runtimes are ok, but heavily dependent on optimization efforts. FLOP counts are much more robust across architectures and independent of the writing effort put into the algorithms.