Aurora on Intel platform. Dawn of the Exaflops Era

There was new information about the Aurora project (the first data about it were published about a year ago, including on Habré ). Let me remind you that, by order of the Argonne Laboratory , a research unit of the US Department of Energy, Intel associates are building a supercomputer with an installed capacity of more than 1 ex-FLOPS. In today's post we will tell what is inside of it - as far as it is known now.



The main mission of the future supercomputer is to carry out calculations for the implementation of a controlled thermonuclear reaction. This task, as everyone knows, is so complex that even a super-device does not seem enough. Nevertheless, it must be solved in order to provide the future humanity with "clean" energy in sufficient quantities. Naturally, Aurora will help in solving other tasks, such as city planning, treatment of serious illnesses and so on.

From a hardware point of view, the supercomputer is made according to the cluster principle; each cluster node contains 2 third-generation Intel Xeon Scalable processors (Sapphire Rapids) and 6 Ponte Vecchio GPUs. The latter is a multi-chip heterogeneous module, assembled using Foveros 3D and EMIB (Embedded Multi-Die Bridge) technologies. In the system, they are connected on a one-to-one basis using a high-speed bus.

The development of Cray - Shasta , which is very flexible in terms of architecture and the types of nodes used, was chosen as the clustering platform . Cray Slingshot technology is responsible for exchanging data within the cluster. As you can see from the diagram, each node has 8 interfaces for connecting to the Slingshot factory.



In total, the supercomputer is equipped with more than 10 petabytes of RAM. Data storage is organized using a distributed asynchronous file system with a maximum capacity of 230 Petabytes and a throughput of 25 TB / s. The photo below shows one Intel Aurora computing unit.

The software basis of the supercomputer is the Intel ONAPI platform , which includes the well-known Intel compilers (C, C ++, Fortran), programming models (DPC ++ / SYCL, OpenMP, OpenCL), libraries (oneMKL, oneDNN, oneDAL) and tools (VTune, Advisor and etc.).

And since the Russian Intel teams are actively involved in the development of the ONAPI component, we can safely say that Aurora is a bit and our Russian Dawn :)

The construction of the Intel Aurora supercomputer is scheduled for completion in 2021. Whether he will be the winner in the “ex-flop race” is not yet clear, but one thing is clear: not a single block will remain without work. Yes, and less powerful computers, too, surely there is a case.