New feature: multi-GPU

[TysonRayJones]

contributors: Tyson Jones
release: v4

QuEST will support deployment to multiple GPUs through hybridising distribution and GPU-acceleration. We support both loosely-coupled GPUs (wherein each GPU lives in separate compute nodes connected over a relatively low-speed network), and tightly-coupled GPUs (wherein GPUs share a direct highspeed interconnect, like an NVLink). This accomplishment results from QuEST's new modularised architecture, enabled by the new algorithms of this manuscript.

• users benefits

  • Users can throw more of their supercomputer resources at simulation, increasing parallelisation.
  • Users can make use of the high memory bandwidth of GPUs interconnected with NVIDIA's GPUDirect or AMD's Infinity Architecture for increased parallelisation, or to simulate larger systems.