Nvidia's chips have come a long way from their humble origins. Today they perform hundreds of septillions of matrix multiplications to power leading large language models. This evolution was thanks to a daring business move from Huang, who funneled the profits Nvidia generated from its gaming business to develop a software layer called Compute Unified Device Architecture, or CUDA.
"You can think about it as the programming environment for GPUs," Bernstein analyst Stacy Rasgon said of CUDA, which was introduced in 2006. Using CUDA, developers can customize GPUs for general-purpose tasks. It's the primary reason Nvidia was able to grow into the AI giant it is today, while dozens of competitors died off or were acquired in the early aughts, according to Rasgon. But for many years CUDA was a black hole bogging down Nvidia's gaming business as it burned through cash.
Nvidia's AI bet achieved proof of concept among the scientific community in 2012. A deep-learning model called AlexNet, trained on Nvidia GPUs, wowed the AI community with its record-breaking accuracy in identifying images. Armed with specialized gaming hardware, AI researchers continued to use games as a laboratory for developing advanced algorithms. Throughout the 2010s, DeepMind achieved a series of AI breakthroughs this way, and the lab continues to explore game-driven AI research today.
