Date/Time
Date(s) - 27/10/2021
3:30 pm - 5:00 pm

Speaker: Prof. John L. Gustafson

Topic: Emerging Alternatives to IEEE Floating-Point

Date of Presentation: Wednesday, October 27, 2021

Location:  Online

Recorded Talk: Link to the Recorded Talk

Passcode: Q^f1@NK2

Abstract:

The need for a better way to represent real numbers on computers is being motivated by the needs of Artificial Intelligence (AI) and High-Performance Computing (HPC). Both areas are increasingly limited by operations-per-watt. The IEEE 754 Floating-Point Standard dates from the 1980s when engineering designs were limited by the cost of transistors and not by memory bandwidth and power consumption. The recently-introduced posit number format is being adopted rapidly as a potential replacement (or supplement) for IEEE floats; it has more information-per-bit, allowing computations to achieve higher accuracy with fewer bits of storage. Comparisons will be shown for posits compared to various float formats (including the Google bfloat) for AI and HPC workloads, showing that improvement factors of 2x to 4x are achievable simply by re-thinking the way we represent numbers on computers.