Next Generation Arithmetic for HPC and AI: An Update
September 19, 2018 at 2:00pm (lunch talk)
Posit arithmetic, a form of universal number (unum) computer arithmetic, was introduced at Stanford in February 2016, as a hardware-friendly drop-in replacement for floating-point arithmetic that offers higher speed, accuracy, and dynamic range using the same number of bits. Since that introduction, posits have proved their merits in machine learning and inference for Artificial Intelligence (AI) as well as PDE solvers, linear algebra, and Fourier transforms for High-Performance Computing (HPC). An open-source fast software library, SoftPosit, has been completed that is closely based on Berkeley’s SoftFloat library that allows easy and experimentally fair comparison between posits and floats. A scalable Verilog generator now exists for creating FPGA and VLSI hardware. Dozens of explorations of posit arithmetic advantages are underway at companies both large and small, and at universities around the globe. New results support the thesis that traditional floating-point arithmetic has become obsoleted by posits, and that posits, surprisingly, can also outperform fixed-point arithmetic approaches currently used for AI and signal processing.
Dr. John L. Gustafson is a Professor in the School of Computing at the National University of Singapore and the inventor of several novel forms of computer arithmetic first introduced in his 2015 book, “The End of Error: Unum Computing.” He is best known for his 1988 argument showing that parallel processing performance need not be limited by “Amdahl’s law,” now generally known as Gustafson’s law. Previously, he has been Senior Fellow and Chief Product Architect at AMD and a Director of Intel Labs. He is a recipient of the inaugural Gordon Bell Prize and is a Golden Core member of IEEE.