The world's fastest computers, known as supercomputers, are used for calculations too big for conventional computers. A classic example is making climate calculations. However, there are many more applications, such as automatic design of (parts of) devices, where strength calculations need to be calculated for very many cases. Recently, this has been joined by the new applications in data science and AI. This has created a renewed interest in supercomputing.

The fastest computer, the US Frontier computer, can reach a maximum speed of more than an exaflop: more than 10^18 floating point calculations per second. This speed is achieved by dividing a calculation among hundreds of thousands of processors, which then execute it in parallel. However, the top speed can only be approached if the computation is composed of operations that are completely independent of each other. If not, processors will have to wait for each other until required partial results are available. As a result, a simple operation like adding a row of numbers can become very time-consuming. Creating algorithms that can be efficiently executed in parallel is the art of parallel computing.

The talk will provide an overview of the architecture of current supercomputers. In particular, DelftBlue, TU Delft's new computer, will be discussed. Afterwards, a number of simple operations will illustrate how a supercomputer can be programmed efficiently by performing calculations in a different way. The talk will conclude with a preview.

Speaker: Martin van Gijzen, Professor in High-Performance Computing at Delft University of Technology