Chip promises brain-like AI in your mobile devices

Eyeriss' trick is to avoid swapping data when possible. Each of the cores (which effectively serves as a neuron) has its own memory, and compresses data whenever it leaves. It also keeps the amount of work to a minimum. Nearby cores can talk directly to each other, so they don't need to talk to a central source (say, main memory) if what they need is close at hand. On top of that, a special delegation circuit gives cores as much work as they can handle without going back to fetch data.