A Perspective on Exascale Computing

The first ExaFlop-capable machine will provide an unprecedented computing power of 10^18 floating point operations per second. It will be likely delivered in the 2020-2021 period in US or China with Japan and Europe following them later. Its main design constraint is power consumption that needs to be less than 40 MW, the power provided by a small-size power plant. To meet the constraint of such a “small” power budget, ExaFlop-capable machines will be using technologies providing the highest number of operations per Watt and reducing data movement as much as possible. For this reason, Exascale machines will be necessarily equipped with accelerators, such as GPUs and FPGAs, and heterogenous memories, such as high-bandwidth 3D-stacked and non-volatile memories. While the hardware of an Exascale supercomputer is somehow known at this point in time, it is still unclear if the software, both programming interfaces and scientific applications, will be ready to exploit Exascale systems in 2020-2021.

In this talk, leveraging my experience gained in participating in European Exascale projects, I will discuss the main challenges we are facing in designing and implementing software for using efficiently Exascale machines, a system with roughly a million of computing nodes featuring accelerators and heterogenous memories. I will show that the main Exascale challenge is now a software challenge. I will discuss software requirements imposed by Exascale supercomputer architectures and different programming approaches that might satisfy these requirements.