High-Performance Computing Trends

Abstract

The decade towards 2020 was marked by the target of a supercomputer with 1 EXA(=1018) floating-point operations/s. While this was a bit too ambitious, the 5-year period 2014–2019 saw a 5-times increase in the throughput of the TOP 10 supercomputers with constant electric power, which means a 5-times improvement in energy efficiency. With this jump in efficiency, two of these TOP 10 have also taken the lead among the TOP GREEN supercomputers. 3D integration, closing the processor-memory gap, graphics accelerators and machine-learning with neuromorphic architectures are among the essentials to possibly maintain this rate of improvement in energy efficiency. The intelligent neuromorphic architectures, yet in their early era, have to produce 100× improvements in energy efficiency and in throughput in the new decade, in order to limit and to manage the data explosion in the INTERNET and in the expected autonomous mobile world. The potential INTERNET energy crisis has been updated in 2018 to a required electric power of 928 GW in 2030, 21% of the expected total global need.