John L. Gustafson

John Gustafson has had a long career in high-performance computing.

John was a prodigy, growing up in Iowa. He won his school's science fair with a radio transmitter he had built, when he was only six years old. His high school refused to let him take Physics or Chemistry because he had, just for fun, taken the SAT achievement tests in those subjects when he was in 10th grade, and he had scored a perfect 800 on both of them.

As a Caltech freshman, he discovered that his algorithm for solving linear systems on HP programmable calculators was for sale in the campus bookstore. In his sophomore year, he discovered a new way to compute the gamma function that converged far faster than existing methods, for which he received the Eric Temple Bell award. His flare for numerical analysis allowed him to graduate from Caltech with Honors.

After getting his MS and PhD from Iowa State, he joined a company called Floating Point Systems, where he led an effort to build the world's first massively parallel computer using commodity parts and distributed memory. It scaled up to 16,384 processors, and customers said he was nuts. For one thing, the national laboratories told him, no one is ever going to do supercomputing by filling an entire room with commodity processors; you need a vector mainframe like a CRAY. For another, Amdahl's law says you quickly reach the point of diminishing returns with parallelism, so no system will ever scale beyond about eight processors. John formulated a counterargument, but did not publicize it until years later when he was able to demonstrate thousand-fold speedup on an nCUBE computer with 1024 processors, at Sandia National Laboratories.

His management encouraged him to publish his counterargument, and his 1988 paper, "Reevaluating Amdahl's Law" changed the course of computing and won him the inaugural Gordon Bell Prize, sometimes called "the Nobel prize of supercomputing". His argument, now known as Gustafson's law, is taught alongside Amdahl's law in just about every university course on parallel computing. John's work ushered in widespread acceptance for parallel computing, including multicore processors, GPUs, and huge networked data centers. Every time you do a web search, play a video game, or use a smart phone, you are using technology that John made possible.

John returned to Iowa State to found the Scalable Computing Laboratory at the Department of Energy's Ames Lab. In the mid-1990s, a project formed to reconstruct the lost Atanasoff-Berry Computer, and John joined that team, eventually becoming its leader and completing the project. For that accomplishment, he was awarded the IEEE Computer Society's highest honor, Golden Core Member.

Since then he has been a Director of Research at Intel and a Senior Fellow at AMD as Chief Product Architect. But he wanted to lead another revolution. this time to redefine how computers do arithmetic, and he wanted it to be completely open-source. So he gave himself a sabbatical from private employers and wrote the book, The End of Error: Unum Computing. That book quickly became a best-seller in its category and attracted the attention of the Singapore government. Singapore invited him to come to that country as a full Professor at National University of Singapore and do research at their government lab, A*STAR. He spent six years there and made the breakthrough of finding a drop-in replacement for standard floating-point arithmetic that gives better answers with fewer bits, called "posit" arithmetic. His second book is Every Bit Counts: Posit Computing, and once again it appears John has created a watershed in computing. The original paper on posit computing has over 500 citations and the idea has become available commercially, so this is not simply an academic exercise. The new format has proved particularly adept at AI (both machine learning and inference), producing better results with less energy consumption.