The DEEP/-ER architecture: a modular approach to extreme-scale computing.

Accelerators arrived to HPC when the power bill for achieving Flop performance with traditional, homogeneous systems became too expensive. By attaching graphic cards (or many-core processors) to general purpose processors, certain application kernels could be sped up at a much lower Watt/flop rate.In 2011 the DEEP project recognised the advantages that heterogeneous processors offered to HPC, but raised the question whether the host-device approach chosen until then for their integration was the most beneficial. The proposed alternative was based on the construction of a cluster of autonomous accelerators, so-called "Booster". The "Cluster" and the "Booster" in DEEP are nothing else than two compute modules with different hardware designs, addressing each different application requirements. By connecting each other with a high-speed network and providing a full software stack that supports them jointly, the resulting "Modular Supercomputer" behaves as a single machine in which users can freely choose resources from any of the compute modules.I will present the architecture and concepts realised in the DEEP and DEEP-ER projects, some of which are going into production this year within JSC's JURECA system. A glimpse of the further developments planed within the just started DEEP-EST project, will also be presented