Nvidia’s director of business development for the life-sciences industry, Mark Berger, was responsible for expanding the use of GPUs in chemistry, biology, and materials science. He followed much the same playbook that Oliver Baltuch had used when trying to raise Nvidia’s profile among its prospective partners in the tech industry, as we saw in chapter 6.

First, he gave away GPUs to researchers and informed them about Nvidia’s substantial investments in creating basic software libraries and tools for CUDA. Although the company might not have been familiar with the esoteric computational problems scientific users might perform, it recognized that those users would rather spend their time designing experiments than building the foundational math libraries they all needed. As a result, the developer tools Berger provided alongside the cards themselves made adoption of CUDA much faster—and helped him establish strong relationships with scientists.

PC games—in particular, first-person shooters—could increasingly produce realistic simulations of physics. When they used GPU processing in its traditional, graphics-acceleration role, these games could calculate the path of a bullet, from the moment it was fired from a gun to the effect of wind on its trajectory to the spalling it produced when it hit a concrete wall. All of these applications relied on various permutations of matrix multiplication—the same math used to solve complex scientific problems.

But because Kirk’s was the first course of its kind, there was no common syllabus or set of standards, no textbook to use. So Kirk and Hwu wrote one. Their first edition of Programming Massively Parallel Processors, which was published in 2010, sold tens of thousands of copies, was translated into several languages, and was eventually used by hundreds of schools. It was a major inflection point in attracting attention, and talent, to CUDA.

Walker later joined GlaxoSmithKline, the pharmaceutical and biotechnology company, as head of scientific computing. The first thing he did was build a data center cluster by using thousands of GeForce gaming cards that cost only about $800 apiece. This caught the attention of Nvidia’s vice president of health care, Kimberly Powell, who called Walker and said, “You’re at GSK now. You need to be buying our enterprise products.” “No,” countered Walker. “I should be doing what’s best for my employer. That’s my job.”

He became increasingly frustrated with Nvidia’s attempt to wring more money out of him, when he had done so much to make CUDA more than just a niche product for well-resourced developers and academics. The architecture wouldn’t have been as successful if Nvidia had restricted its use to cards that cost thousands of dollars; it would have been almost as expensive to use CUDA as it would have been to design a custom ASIC.