Abstract: High-performance computing hardware is continuously evolving, and one of the fundamental aspects which has relatively recently started to change is the mathematical hardware, the introduction of the IEEE 754 floating-point standard. Increasingly, hardware for 16- and 8-bit floating-point computations is included, mainly driven by the market’s need to accelerate operations that applications that utilize techniques from machine learning rely on. These arithmetics have high throughput, but they present the users with non-standard numerical behaviours, mixed-precision operations (input and output formats differ), novel rounding modes, as well as require the development of modified or completely new algorithms for use in applications, which now go beyond machine learning and into areas such as climate and weather simulations or radio astronomy.

In this talk we will discuss low-precision number formats, matrix multiply-accumulate hardware that is increasingly appearing in the datacentre GPUs, non-standard rounding routines, progress on testing the hardware to determine various numerical features, as well as currently known effects on numerical algorithms that utilize this hardware.

Bio: Mantas Mikaitis received a B.Sc. (Hons.) degree in Computer Science and a PhD degree in Computer Science at the University of Manchester, Manchester, United Kingdom. He is currently a Lecturer at the School of Computing of University of Leeds, Leeds, United Kingdom.