Why is the chip that drew pictures now doing the maths?

That time, that place

A computer's central processing unit, the CPU, is a few clever workers, handling complex decisions quickly, one after another.

A graphics chip — a GPU — is different. It must compute a colour for each of two million pixels, and each computation is simple. So it holds thousands of simple workers and sets them all going at once.

Why this question

In the early 2000s a few researchers tried something strange: asking the graphics chip to compute things other than pictures.

The method was grotesque. You had to disguise your equation as a drawing problem — packaging data as if it were a texture image — before the chip would take it.

But it was fast. Very fast.

In 2007 Nvidia removed the disguise, releasing tools that let you hand the chip a calculation directly.

What was found

This matters because the arithmetic of a neural network is exactly the kind a graphics chip is good at.

Training is repeated multiplication and addition across enormous tables of numbers. Each multiplication is independent of the others, so they can all be done at once. Vast numbers of simple operations in parallel — what the graphics chip was born doing.

The 2012 explosion in deep learning happened because these tools existed. And the racks now filling data centres, and the electricity question that has become national policy, start here.

A chip built to draw game frames became the heart of artificial intelligence.