AMD's Vision: AI's Demand for Yottaflop-Level Computing Power

The future of artificial intelligence is set for unprecedented levels of computing power, epitomized by a term unfamiliar to most: the yottaflop.

In her address at CES 2026, AMD's CEO, Lisa Su, projected that to accommodate the explosive growth of AI, the world will require computing capabilities exceeding '10 yottaflops' within the next half-decade.

Encouraging the audience to engage, Su inquired if anyone knew the meaning of a yottaflop. When no one responded, she elucidated that a yottaflop represents a one followed by 24 zeros, signifying 10,000 times the computational power available in 2022.

To put it into perspective, a flop is a standard measurement for a basic arithmetic task. A gigaflop corresponds to a billion such operations per second, whereas a yottaflop equates to a septillion operations per second.

Experts theorize that with 10 yottaflops, we could simulate entire planetary systems down to their atomic intricacies.

Current and Future Computational Benchmarks

In 2022, the combined computational capacity dedicated to AI was approximately a zettaflop, defined as a one followed by 21 zeros. By 2025, this had surged beyond a hundred zettaflops, as noted by Su.

This leap in computational demands for AI is unparalleled, akin to technological evolutions unseen in the computing domain, Su remarked at the conference held in Las Vegas.

This prediction hints at a system 5.6 million times faster than today's leading supercomputer, El Capitan, managed by the US Department of Energy.

Challenges in Scaling Up AI's Computing Needs

The existing AI computational requirements already strain America’s energy infrastructure. The development of supporting energy systems is recognized as a significant hurdle in scaling up AI capabilities.

During her keynote, Su took the opportunity to introduce AMD's upcoming AI chip designs, notably the MI455 GPU, as the company intensifies its foray into providing data-center technology for clients like OpenAI.