Note to self: LHS is basically fixed so if we want improve performance, we should strive to decrease energy/op

One important detail here is the fact that Dennard's scaling has essentially halted. Specifically, Dennard's scaling is the principle underpinning Moore's Law, and stated that adding more transistors to a chip would not change the power density; in other words, we could add significantly more transistors to a chip, but individually they would require less power, allowing the development of faster computers at the same power requirement. Sadly, the magic of Dennard's scaling has ended with the breakdown of Moore's Law due to the fact we cannot shrink transistors forever. This is basically why we now have to work much more with the 2nd term (energy/op) rather than the first.

Also good to note that the goal of this slide is to work around a fixed power esp for use in mobile devices/ edge devices

Assume power is fixed. We need to make more energy efficient algorithms, or reduce the number of joules for every operation we have. I wonder what some ways to do that are.

How does this work if different ops consume different amounts of power, such as AVX2?