The actual workload in ML model training is totally deferrable. It ought to be very compatible with intermittent power, and actually great for the renewable grid. The two solutions for intermittence are: buy batteries or buy so many solar panels that you still get enough power even when they are operating at like 30% efficiency. A workload that will pay big bucks for any excess power you’ve got will provide the needed ROI for the latter solution.
Unfortunately the chips are all to expensive so everybody wants to run them full-out all the time.
Unfortunately the chips are all to expensive so everybody wants to run them full-out all the time.