Special-purpose chip that performs some simple, analog computations in memory reduces the energy consumption of binary-weight neural networks by up to 95 percent while speeding them up as much as sevenfold.