Hank Elsner Live Wire
Tech, crypto and space news, read live while the laser runs. · RSS
Home / Deep Dive · · By Hank Elsner · about 4 min read aloud

How Video Game Chips Accidentally Took Over AI

A chip built to draw explosions in video games turned out to be perfect for teaching computers to see. That accident built Nvidia.

How Video Game Chips Accidentally Took Over AI

The most valuable chip company in the world got there because of video games. Not on purpose. By accident. Here's how a chip built to draw explosions in a shooter ended up running every AI on Earth.

Two kinds of chips

Your computer has a CPU, the main processor. Think of it as one brilliant professor. It can do anything, really fast, but mostly one thing at a time, or a handful.

A graphics card has a GPU. Think of it as a thousand students. Each one is a lot simpler than the professor. But give them all a simple math problem at the same time, and they crush the professor on sheer volume.

Why games needed the students

Look at your screen. Millions of tiny dots, pixels. In a video game, every one of those pixels has to have its color worked out, sixty or more times every second. That's millions of small, similar calculations, all at once. Perfect for a thousand students, terrible for one professor.

So through the 1990s and 2000s, companies like Nvidia raced to build graphics chips with more and more of those little cores, because gamers wanted better graphics.

The side door

In 2007 or so, Nvidia released a software tool called CUDA. It let programmers use those graphics chips for any kind of math, not just drawing pictures. At the time, it seemed like a niche thing for scientists.

Then, in 2012, three researchers at the University of Toronto, Alex Krizhevsky, Ilya Sutskever and their adviser Geoffrey Hinton, entered an image recognition contest called ImageNet. The goal: teach a computer to look at a photo and say what's in it. Their program, called AlexNet, crushed the competition. And they trained it on two Nvidia gaming graphics cards.

That was the moment the AI world woke up. Sutskever later co-founded OpenAI. Hinton later shared a Nobel Prize in physics for foundational work on these kinds of systems.

Why AI loves the students

Here's why it worked. Modern AI, under the hood, is mostly one kind of math: multiplying huge grids of numbers together, called matrices. Over and over. Billions of times.

That's exactly the kind of job you want to split across a thousand students. Each one handles a small piece of the grid at the same time. The same trick that colors pixels in a game turns out to be the trick that makes a computer learn.

From gaming to everything

After that, Nvidia leaned all the way in. It built chips designed specifically for AI, then whole cabinets of them wired together, then whole rooms. Remember the Vera Rubin system from tonight's news? That's the great-great-grandchild of a gaming card. Same idea, a thousand students, scaled up into a building.

And because Nvidia had CUDA years before anyone else, most of the world's AI software was written for its chips. That head start is a big part of why Nvidia got so far ahead. It's also why the Anthropic claim about moving work to AMD chips in days was such a big deal: the software has always been the moat.

The lesson

Here's what I love about this story. Nobody at a gaming company in 1999 thought, "We're building the brains of AI." They were trying to make better explosions. Sometimes the most important thing you build is the thing you built for a totally different reason. Keep tinkering.

Ask the chatWhat was the first graphics card or game console you ever owned? Type it.
On the laserCard: "Built for games. Took over the world." For the oldest graphics card in chat.