THERE’S AN APOCRYPHAL story about how NVIDIA pivoted from games and graphics hardware to dominate AI chips – and it involves cats. Back in 2010, B

NVIDIA and the battle for the future of AI chips

submited by
Style Pass
2021-06-20 20:00:10

THERE’S AN APOCRYPHAL story about how NVIDIA pivoted from games and graphics hardware to dominate AI chips – and it involves cats. Back in 2010, Bill Dally, now chief scientist at NVIDIA, was having breakfast with a former colleague from Stanford University, the computer scientist Andrew Ng, who was working on a project with Google. “He was trying to find cats on the internet – he didn’t put it that way, but that’s what he was doing,” Dally says.

Ng was working at the Google X lab on a project to build a neural network that could learn on its own. The neural network was shown ten million YouTube videos and learned how to pick out human faces, bodies and cats – but to do so accurately, the system required thousands of CPUs (central processing units), the workhorse processors that power computers. “I said, ‘I bet we could do it with just a few GPUs,’” Dally says. GPUs (graphics processing units) are specialised for more intense workloads such as 3D rendering – and that makes them better than CPUs at powering AI.

Dally turned to Bryan Catanzaro, who now leads deep learning research at NVIDIA, to make it happen. And he did – with just 12 GPUs – proving that the parallel processing offered by GPUs was faster and more efficient at training Ng’s cat-recognition model than CPUs.

Leave a Comment