At last, my friends, my mother, and the rest of the world get to learn what I’ve been working on—in secret—since I came to Micron exactly three years ago. At the venerable Supercomputing Conference this week, Micron announced a technology that is extremely rare—even to experts immersed in the exotic world of the highest-performance computing systems on the planet. It’s a completely new way to compute, and I’ll tell you why.
Serial processors improve by small increments from year to year, but the big computing challenges that the experts care about most can only be solved by harnessing the power of several processors working together. And Micron's new Automata Processor does just that; and it happens to be the most massively parallel compute engine ever devised.
Inventions never hatch fully formed from the brain of the inventor (at least I don't know of any), and the Automata Processor is no different. It started with a great idea, but it took the unique synergy of our team and a few fortuitous right and wrong turns to get us where we are today. It was less than two years ago when I realized that what we were onto was not just good, but that it would most likely cause every computer science textbook to be rewritten—and that it would completely change the way we think about and analyze human-constructed parallel computing systems.
We’ve already been working on the Automata Processor for several years, but we still have a long path ahead. We’ve started to work with high-performance computing experts, but we will eventually work with experts in all computing technologies—down to small mobile devices—to build a plethora of applications. We’ve also begun the long journey to perfect our chip. While we’ve already seen our very first Automata Processor chip do remarkable things, we can only guess (often incorrectly) about how application designers will want to use it; we’re gaining more insight about customer needs to enhance our design. We’ve also engaged with computer scientists to discuss Automata’s processing and parallel computing. We’ve found that, as with every huge change, this one will bring some disruption to the entire ecosystem—including the academic segment.
For example, I had the opportunity to talk to one of the most eminent computer scientists in the world—the expert that taught many of us some of the science behind the Automata Processor (the theory of automata). This professor first asserted that the Automata Processor could not possibly exist. After I described in some detail how our Automata Processor worked, I won the professor’s grudging acknowledgement that such a chip could be built; but the professor did not believe that Micron, or a company like us, could build such an innovation in the foreseeable future. He made the association to the great computer scientist John von Neumann's idea for a gold-mining robot sea monster.
This so-called monster would theoretically swim in the ocean, filtering the water for the 1 in 10 million parts of gold content. When the robot accumulated enough gold, it would return to shore and replicate itself using other material that it filtered from the sea. The process would repeat, and the number of sea monsters would continue to increase, until they extracted all the gold from the ocean.
Ok, so the sea monsters may be far-fetched, but our Automata Processor does have a slightly fantastical nature. It implements nondeterministic finite automata to compute with astonishing density. A conventional compute engine requires exponentially larger resources to do that. There is only one other computing device that I can think of that can compute at that magnitude—it’s called the brain! A computer scientist at Stanford estimated that a computer system with the computational capabilities of the human brain, weighing 3 pounds and consuming about 20 watts, would consume a whopping 10 million watts if constructed.
The Automata Processor, which is composed of massive numbers of interconnected but wholly autonomous elements, computes much more like the brain than a conventional processor. Our scientist friend John von Neumann, also considered the father of conventional processor architecture, thought deeply about the computational model of the brain when working on the manuscript, "The Computer and the Brain,"—a piece of work that has greatly inspired our team. Von Neumann's robot sea monster is not just a computer scientist's tongue-in-cheek fable; it’s a metaphor for the power of driving effective parallelism with a simple organizing concept. The Automata Processor, in fact, excels at finding the one part among 10 million others and is readily replicable—just like von Neumann’s robot that mined gold from seawater.