>Nothing crazy about it. TPU-like stuff is ~10x the energy efficiency of GPUs and several times the speed. When you're spending megawatt-hours and days to train

>Nothing crazy about it. TPU-like stuff is ~10x the energy efficiency of GPUs and several times the speed. When you're spending megawatt-hours and days to train a single model, it adds up in both real and opportunity costs. could you share the stats on this? Google told me to use a K-80 for training.