50 Spine-Tingling Future Predictions by Jensen Huang

CHAPTER 01 — The structure of computing shifts from chips to systems

The CPU era ends and the GPU will seize sovereignty over computation

In 2016, the logic of the computer industry was simple. It was believed that most performance gains came from making the CPU faster. A faster CPU, more memory, a more efficient operating system — that was enough. The dominant belief was that as long as Intel upheld Moore's Law, the world would keep getting faster. Data centers were packed with Intel chips, and that structure had not changed much in decades.

But Jensen Huang saw it differently. Deep learning, an AI learning method that finds patterns in data, must process hundreds of millions of simple operations simultaneously. The CPU is built to process a single operation very fast, but it was not suited to such parallel computation. By analogy, if the CPU is a handful of senior officials handling complex administrative tasks, the GPU is closer to tens of thousands of skilled workers performing simple tasks all at once. He saw a fundamental limit in scaling AI computation with a CPU-centered structure, and predicted that sovereignty over computation would inevitably move to the GPU. The GPU, designed for game graphics, fit this structure better, because the way thousands of small cores operate simultaneously dovetailed with the nature of AI training.