Over the last decade, specialized heterogeneous hardware designs ranging from Cell over GPGPU to Intel
Xeon Phi have become a viable option in High Performance Computing – mostly due to the fact that these
heterogeneous architectures allow for a better flops-per-watt ratio than conventional multi-core designs.
However, the corresponding programming models for heterogeneous architectures so far remain limited.
Neither offload models (like Cuda from Nvidia, OpenACC or Intel offload directives) nor the native
execution on the accelerator (e.g execution on Intel Xeon Phi) is able to provide a single cohesive view of
the underlying fragmented heterogeneous memory.
The upcoming new GASPI standard will able to bridge this gap in the sense that GASPI can provide
partitioned global address spaces (so called segments), which span across both the memory of the Host
and e.g. an Intel Xeon Phi.
Download the solution brief here : Fraunhofer ITWM demonstrates GPI 2.0 with Mellanox Connect-IB™ and Intel® Xeon Phi™