Nvidia's got a cunning plan to keep powering the AI revolution

Nvidia’s AI journey started at Joanie’s Cafe in Palo Alto, California, in 2010. “We got into AI after a breakfast meeting I had with Andrew Ng,” says Dally. At that breakfast, Ng, a well known AI researcher who was working with Google Brain at the time, explained how Google was training AI systems to recognise photos of cats with the help of 16,000 central processing units, or CPUs. After pointing out next to no one has 16,000 CPUs at their disposal, Dally laid down a challenge. “I bet we could do this with way fewer GPUs,” he said to Ng.

Shortly after the encounter, Dally asked Nvidia researcher Bryan Catanzaro to team up with Ng and do just that. “We achieved roughly the same performance with 48 GPUs that they had done with 16,000 CPUs,” says Dally. “At the time it was really clear to me that this was going to transform everything.”

From a bet about recognising cats, Nvidia’s AI business has exploded. The company doesn’t disclose exactly how much it is currently investing into the field, but filings reveal it spent a total of $1.8 billion (£1.4b) on all research in the last financial year. Key areas of investment were gaming, AI and automotive. The effectiveness of its GPUs for artificial intelligence projects has created a scramble amongst Nvidia’s competitors, with Intel, Google and even Facebook investing huge sums of money to try and catch up.

In July 2018, Google announced it was developing its own AI chips for on-device machine learning. Google’s bet is on a new kind of chip: a tensor processing unit designed specifically for neural network machine learning. And a separate, equally important tussle is also unfolding. As well as hardware, Nvidia’s code library is also going up against rival frameworks such as Google Tensorflow or Facebook PyTorch. Think of it as deep learning as a service. And, against all the odds, Nvidia is still holding its own.

The reason? As the cat challenge showed, GPUs are brilliant at handling AI tasks. That’s because GPUs are better than CPUs at running the parallel computing problems associated with AI, where many calculations are carried out simultaneously. Parallel computing is particularly prevalent in neural networks because they are designed to work in a similar way to animal brains, which can do many tasks at the same time, or in parallel. Eventually, Nvidia wants to be the one building the artificial brains inside every phone, computer, robot and autonomous car.

Led by CEO Jen-Hsun Huang, the company effectively preempted the AI boom back in 2005 when it created software that allowed its GPUs to process the millions of minuscule computations that would eventually be required by modern AI. Its GPUs are now in data centres around the world, powering AI tasks for thousands of businesses. Between 2014 and 2018, the company was boosted by a 524 per cent increase in revenue for its data centre-optimised chips.

“We’re the supplier to the whole world,” says Dally. “Everyone’s training their deep neural networks on Nvidia GPUs,” he continues, adding that the company’s T4 chip is widely used for inference tasks. Facebook and Google use Nvidia’s platform to power the AI features on their platforms. Almost all the driverless car companies are using Nvidia technology. The company has also signed deals with Tencent, Alibaba, and Baidu.

As with the robot in the kitchen, Nvidia’s other AI experiments appear similarly eclectic and oddball. It’s built a creative neural network, known as a generative adversarial network, or GAN, that’s scarily good at creating hyper-realistic images of human faces. In a paper released in December 2018, the company showed just how clear its GAN faces are compared to GAN faces from four years ago.

Faces aren’t the only thing that Nvidia’s GANs can create, or fake. Nvidia has also developed a translation network that can turn photos taken in winter into beautiful summery scenes. Got a rainy photo of your wedding you want brightening up? Nvidia’s AI can make it look just like the sun was shining all along. And it doesn’t stop there. Nvidia showed off the world’s first video game demo to use AI-generated graphics at the end of last year. The demo was a driving simulator that was built by combining AI-generated visuals with a standard video game engine.

But there have been bumps in the road. And they hint at a rougher ride to come. At the start of the year, Nvidia was forced to cut its fourth quarter revenue forecast by $500 million in a move that saw its stock tumble nearly 18 per cent. The cryptocurrency crash that rumbled through much of 2018 was extremely painful for Nvidia’s bottom line, costing it an estimated $23 billion in market value. But it was also a blessing – albeit an expensive one.

For the time being, the adaptability of Nvidia’s GPUs give it an edge over its rivals: they’re integral to data centres, self-driving cars, research labs, cryptocurrency mining rigs – you name it, there’s a GPU in it. But that could soon change. And that’s where Nvidia’s research labs come into play.

“There’s a slew of new AI hardware startups,” says Stephen Merity, an AI research consultant and the founder of a startup in stealth mode. “They may well be a threat but it’s a huge open question. I can’t imagine any of those startups hitting mass scale early on.”

One such startup is UK-based Graphcore, which is now valued at over $1 billion (£760,000) thanks to funding from the likes of BMW, Sequoia, and the founder of Google-owned DeepMind. Graphcore claims its intelligence processing units, which are yet to go into mass production, are 100 times faster than any existing systems. For Nvidia, that could create a financial headache bigger than the cryptocrash.