Not sure if it fits but there was project halide a while ago which balanced caching calculation in L2 cache vs calculating them afresh.

Examples were given for image processing. This sounded very impressive but I’ve wondered what has happened to it now there’s Machine Learning tools.