Yea the benefit is not going to come in terms of performance for a given model, but in terms of ability to be efficiently quantized.