It's not really any different from the current state of models, it's just that moving the same constraints to a different hardware foundation opens up significant performance gains.