4 ms·

Maybe, but I'd rather pay for consistent access to state of the art quality even if it's slower (which hasn't mattered much while parallelizing)