Comment by tolugenius

I wonder what would "compute" mean if cpus were more efficient at matrix multiplication say 15 years ago. And on the flipside, what it would take to say train a frontier model entirely on cpus in the future.