You can shuffle it back and forth between disk and memory. It's slow, but it works. There are people working on compute-in-memory hardware, like flash chips th

You can shuffle it back and forth between disk and memory. It's slow, but it works. There are people working on compute-in-memory hardware, like flash chips that can do matrix multiplication in place. None of it is close to reaching the market but there's a lot more interest now that neural networks are obviously useful.