Cuda

In the year 2000 Ian Buck wanted to do something that seemed impossible: play Quake III in 8K resolution. Young Buck was studying computer science at Stanford, specializing in computer graphics, and then a crazy idea occurred to him: put together 32 GeForce graphics cards and render Quake III on eight strategically placed projectors. “That,” he explained years later, “was beautiful.” Buck told that story in ‘The Thining Machine’, the essay published by Stephen Witt in 2025 that traces the history of NVIDIA. And of course one of the fundamental parts of that story is the origin of CUDA, the architecture that AI developers have turned into a gem and that has allowed the company to boost and become the most important in the world by market capitalization. And it all started with Quake III. The GPU as a home supercomputer That, of course, was just a fun experiment, but for Buck it was a revelation, because there he discovered that perhaps specialized graphics chips (GPUs) could do more than draw triangles and render Quake frames. In 2006 the GeForce 8800 GTS (and its higher version, the GTX) began the CUDA era. To find out, he delved into the technical aspects of NVIDIA graphics processors and began researching their possibilities as part of his Stanford PhD. He gathered a small group of researchers and, with a grant from DARPA (Defense Advanced Research Projects Agency), began working on an open source programming language that he called Brook. That language allowed something amazing: making graphics cards become home supercomputers. Buck demonstrated that GPUs, theoretically dedicated to working with graphics, could solve general-purpose problems, and also do so by taking advantage of the parallelism offered by those chips. Thus, while one part of the chip illuminated triangle A, another was already rasterizing triangle B and another writing triangle C in memory. It wasn’t exactly the same as today’s data parallelism, but it still offered amazing computing power, far superior to any CPU of the time. That specialized language ended up becoming a paper called ‘Brook for GPUs: stream computing on graphics hardware‘. Suddenly parallel computing was available to anyone, and although that project barely received public coverage, it became something that one person knew was important. That person was Jensen Huang. Shortly after publishing that study, the founder of NVIDIA met with Buck and signed him on the spot. He realized that this capacity of graphics processors could and should be exploited, and began to dedicate more and more resources to it. CUDA is born When Silicon Graphics collapsed in 2005 – due to NVIDIA that was intractable in workstations – many of its employees ended up working for the company. 1,200 of them in fact went directly to the R&D division, and one of the big projects of that division was precisely to take forward this capacity of these cards. John Nickolls / Ian Buck. As soon as he arrived at NVIDIA, Ian Buck began working with John Nickolswho before working for the firm had tried—unsuccessfully—to get ahead of the future with his commitment to parallel computing. That attempt failed, but together with Buck and some other engineers he launched a project to which NVIDIA preferred to give a somewhat confusing name. He called it Compute Unified Domain Architecture. CUDA was born. Work on CUDA progressed rapidly and NVIDIA released the first version of this technology in November 2006. That software was free, but it was only compatible with NVIDIA hardware. And as often happens with many revolutions, CUDA took a while to gel. In 2007 the software platform was downloaded 13,000 times: the hundreds of millions of NVIDIA graphics users only wanted them for gaming, and it remained that way for a long time. Programming to take advantage of CUDA was difficult, and Those first times were very difficult for this projectwhich consumed a lot of talent and finances at NVIDIA without seeing any real benefits. In fact, the first uses of CUDA had nothing to do with artificial intelligence because artificial intelligence was barely talked about at the time. Those who took advantage of this technology were scientific departments, and only years later would the revolution that this technology could cause take shape. A late (but deserved) success In fact, Buck himself pointed this out in a 2012 interview with Tom’s Hardware in 2012. When the interviewer asked him what future uses he saw for the GPGPU technology offered by CUDA in the future, he gave some examples. He talked about companies that were using CUDA to design next-generation clothes or cars, but he added something important: “In the future, we will continue to see opportunities in personal media, such as sorting and searching photos based on image content, i.e. faces, location, etc., which is a very computationally intensive operation.” Here Buck knew what he was talking about, although he did not imagine that this would be the beginning of the true CUDA revolution. In 2012 two young doctoral students named Alex Krizhevsky and Ilya Sutskever They developed a project under the guidance of their supervisor, Geoffrey Hinton. The Nvidia Way: Jensen Huang and the Making of a Tech Giant (English Edition) The price could vary. We earn commission from these links That project was none other than AlexNetthe software that allowed images to be classified automatically and which until then had been a useless challenge due to the cost of the computing it required. It was then that these academics trained a neural network with NVIDIA graphics cards and CUDA software. Suddenly AI and CUDA were starting to make sense. The rest, as they say, it’s history. In Xataka | We can forget about AI without hallucinations for now. NVIDIA CEO explains why