It would be great if these economists and other non-practioners would consult any legitimate domain expert before writing these dumb columns. David Patterson ha

It would be great if these economists and other non-practioners would consult any legitimate domain expert before writing these dumb columns. David Patterson has been writing and presenting on this topic a lot in the last 2-3 years because some widely-cited paper overestimates the cost of training and inference by between 50x and 1000000x depending on how you look at it. His deck is pretty interesting: https://chips-compilers-mlsys-22.github.io/assets/slides/10%20Lessons%204%20TPU%20gens%20+%20CO2e%2045%20minutes.pdf#page=66 https://chips-compilers-mlsys-22.github.io/assets/slides/10%...