Ian Buck built a gaming rig out of 32 GeForce graphics cards while he was a computer science PhD student at Stanford. He first meant it for Quake and Doom. Then his interest turned to a different question: could those chips do something besides draw pictures?
The answer became CUDA, Nvidia’s platform for running general-purpose work on graphics processors. Nvidia calls it proprietary, and it works only on Nvidia GPUs.
From Brook to CUDA
Buck had been hooked on GPUs since his undergraduate years at Princeton, where video games drew him in, and he interned at Nvidia after graduating. At Stanford he wrote Brook, a programming language for general-purpose computing on GPUs. Nvidia and the Defense Advanced Research Projects Agency (DARPA) both backed the work.
In 2004 Nvidia hired him and paired him with John Nickolls, then director of architecture for GPU computing. Together they set about turning Brook into CUDA. The company introduced it in 2006, according to the Computer History Museum, and the first SDK went public on February 15, 2007 for Windows and Linux.
The name started out as Compute Unified Device Architecture. Nvidia later dropped that meaning and rarely spells it out now.
What it actually does
Think of CUDA as two things at once. One is a software layer that manages data and gives programs direct access to the GPU and the CPU as needed. The other is a library of programming interfaces for running calculations in parallel. Drivers, runtime kernels, compilers, libraries and developer tools come in the package.
Before CUDA, a programmer who wanted to use a GPU had to work through interfaces like Direct3D and OpenGL, which demand advanced graphics skills. CUDA is written in C but works with C++, Fortran, Python and Julia, so specialists in parallel programming could use GPUs without becoming graphics experts. It also supports frameworks including OpenMP, OpenACC and OpenCL.
By 2012, GPUs had grown into highly parallel multi-core systems that handle big blocks of data efficiently. For algorithms that chew through large blocks of data in parallel, that design beats a general-purpose CPU. Developers have used CUDA to speed up applications in computational biology and cryptography by an order of magnitude or more, and in game physics such as debris, smoke, fire and fluids.
The AI turn
In the mid-2000s Nvidia saw that its GPUs could serve scientific computing. Universities and national laboratories soon used the chips to power some of the world’s fastest supercomputers.
Then came 2012. Nvidia GPUs powered the AlexNet neural network that year, and the company pivoted again toward AI. By 2015, CUDA development leaned toward machine learning and neural network workloads. The Computer History Museum says most AI models today, from large language models to autonomous driving systems, are trained and run on Nvidia technology.
Jensen Huang founded Nvidia in 1993 with Chris Malachowsky and Curtis Priem, aiming to bring advanced 3D graphics to gaming and multimedia. Its 1999 GeForce 256 was billed as the first GPU. Others have tried to run CUDA code on rival hardware. Projects such as ZLUDA and SCALE target AMD chips, and ZLUDA’s developer was contracted by Intel in 2021 and AMD in 2022 to build it, though neither company released it officially.
