The secret to Nvidia's research success: Failing often and quickly

Yahoo

20-03-2025

In the span of just a few short years, Nvidia (NVDA) has become one of the most important chip companies in the world. Revenues have skyrocketed from $27 billion in the company's fiscal 2023 to $130.5 billion in its fiscal 2025. Share prices are also soaring more than 680% since January 2023.
Not quite a household name like other Big Tech firms, Nvidia is at the center of the global AI push thanks to its powerful chips, including the Blackwell Ultra, which the company showed off during its annual GTC event on Monday.
A number of technologies behind those processors, the ones that power gaming PCs around the world, and the software that runs both all originated in Nvidia's relatively small research and development department: The appropriately named Nvidia Research.
Established in 2006, the group is responsible for everything from Nvidia's ray-tracing technology, which creates realistic lighting for gamers and professional designers, to NVLink and NVSwitch, which allow graphics chips and central processing units (CPUs) to communicate at the kind of speeds needed for AI systems.
Currently, the organization is working on new chip architectures, quantum computing, and software simulators that teach robots and self-driving cars how to navigate the real world.
It's all meant to keep pushing Nvidia forward at a time when it's already riding high. And to do that the research team has adopted a willingness to fail more often than not while also giving promising projects the time they need to succeed, no matter how long it takes.
'We have to realize that most things we start in Research fail, and that's actually a good thing,' explained Bill Dally, senior vice president of research and chief scientist at Nvidia. 'I tell people, you know, if everything you do succeeds, you're not swinging for the fences. You're bunting.'
While Nvidia has developed a number of impressive technologies over the years, the company's research team isn't nearly as large as some of those at other Silicon Valley companies.
'We're a tiny fraction of the size of competitive research labs,' Dally said. 'We're 300 [people] and yet, I think in things that matter, we punch well above our weight. And I think the real measure of that, for me, is our impact over the years on getting things to [a marketable] product.'
According to Dally, the best researchers are those who come up with an idea, test it, and, if it doesn't work out, abandon it without wasting resources on it.
But if a concept looks like it could pan out, the company will continue to chip away at it until it's a worthwhile product or technology.
Nvidia's ray tracing is a perfect example. The product took 10 years to develop but is now used across hundreds of major games and in design software.
'I think it's quite extraordinary that the company was able to follow through on a vision that took more than 10 years to implement,' said Bryan Catanzaro, vice president of applied deep learning research at Nvidia.
'AI is the most important example of that,' Catanzaro, who joined Nvidia as an intern in 2008, explained.
'AI in 2011 was considered old and dumb and dead. It's like, people have been trying this since the 1950s and it's never worked, so why would it work now? But there were a few of us that believed this was really an opportunity and so the company gave us the space to continue trying things out and then to produce kind of incrementally better results, which then led to more investment incrementally,' added Catanzaro.
Nvidia's DLSS, or deep learning super sampling, is another example of a product the company continued to pursue despite early struggles. Introduced in 2019, the first iteration of DLSS improves a game's image quality and performance using AI. But the software didn't hit the mark out of the gate. I remember trying it out on my own computer and not seeing much improvement while playing games.
Fast-forward to today, and the company now offers DLSS 4, which dramatically improves game visuals for even the most resource-intensive titles, including 'Cyberpunk 2077.'
'DLSS 1.0 was not great, and a lot of people thought that it was a bad idea, this was a bad technology. We believed in it,' Catanzaro said. 'I think Nvidia just has this unshakable belief when it knows something is true about the future, it just keeps banging away at it.'
Not every successful research project ends up as a product that directly generates revenue. However, they can help power sales indirectly by driving GPU sales.
'I'm perfectly happy with people developing … applications for GPUs that broaden the market,' Dally explained.
"Recently, our folks did this thing called Sana, which is this text-to-[image] generative network. And so it doesn't go into a product, but it's still a great success because people outside use it, and therefore it fuels demand for GPUs."
That's ultimately the goal. But the company's newly unveiled Blackwell Ultra and Vera Rubin superchip also come at a time when Nvidia is facing increased competition. AMD is offering up its own AI chips designed to rival Nvidia and the company's customers are developing or deploying specialized AI processors of their own.
There are also market-shaking moves like the release of DeepSeek's R1 AI model, which sent Nvidia's market cap plunging nearly $600 billion in January, and the unpredictability of governmental intervention, including tariffs and export controls, which continue to weigh on the company's stock price.
And with tech companies like Amazon (AMZN), Google (GOOG, GOOGL), Meta (META), and Microsoft (MSFT) set to spend billions on AI infrastructure in the years ahead, Nvidia's research efforts become all the more important as it works to ensure its share of that bounty.
It just needs to keep failing quickly and moving forward.
Email Daniel Howley at [e-mail]. Follow him on Twitter at @DanielHowley.