...machines that spread computations across vast numbers of tiny, low-power chips can operate more like the human brain, which efficiently uses the energy at its disposal.
…the leading internet companies are now training their neural networks with help from another type of chip called a graphics processing unit, or G.P.U. These low-power chips — usually made by Nvidia — were originally designed to render images for games and other software, and they worked hand-in-hand with the chip — usually made by Intel — at the center of a computer. G.P.U.s can process the math required by neural networks far more efficiently than C.P.U.s.
G.P.U.s are the primary vehicles that companies use to teach their neural networks a particular task, but that is only part of the process. Once a neural network is trained for a task, it must perform it, and that requires a different kind of computing power.
After training a speech-recognition algorithm, for example, Microsoft offers it up as an online service, and it actually starts identifying commands that people speak into their smartphones. G.P.U.s are not quite as efficient during this stage of the process. So, many companies are now building chips specifically to do what the other chips have learned.
Google built its own specialty chip, a Tensor Processing Unit, or T.P.U. Nvidia is building a similar chip. And Microsoft has reprogrammed specialized chips from Altera, which was acquired by Intel, so that it too can run neural networks more easily.
The hope is that this new breed of mobile chip can help devices handle more, and more complex, tasks on their own, without calling back to distant data centers: phones recognizing spoken commands without accessing the internet; driverless cars recognizing the world around them with a speed and accuracy that is not possible now.
In other words, a driverless car needs cameras and radar and lasers. But it also needs a brain.