I see it making claims about 10x efficiency, but how is tokens / second / watt? The only machines I've seen with the memory bandwidth to effectively do local inference are Mx arm chips on mac.