By eliminating a hidden inefficiency, computer scientists have come up with a new way to multiply large matrices that’s faster than ever.