Lower precision can also speedup inference because it takes less time to perform calculations with fewer bits.