For training I find wherever I can get the cheapest gpus for my training set, usually aws is the winner. But we operate in 3 clouds. For inference it’s complica

For training I find wherever I can get the cheapest gpus for my training set, usually aws is the winner. But we operate in 3 clouds. For inference it’s complicated. A lot of that is cheaper to do on the right low priority/spot/preemptive machine. Tpus seem to be exceptional at operations like soft max. I wonder how well fpgas might be good at some ml applications you might be thinking of. It all depends honestly.