Architectural requirements for a hybrid GPU/CPU middleware
Burkhard Zink · 2008
The massive parallel performance of graphics processing units (GPUs) has recently inspired increased research activity to employ graphics hardware for scientific and business computation purposes. Current hardware can already perform at peaks of 350 GFlop/s for single-precision operations, and the graphics card manufacturers, most notably AMD and NVIDIA, plan to support double-precision operations in production hardware in the near future, making GPUs an increasingly interesting target for parallel scientific codes.