Hardware Aware Numerics and High Performance Computing

This research field at EMCL aims at developing hardware-aware numerical schemes, platform-specific implementations and simulation software for research with respect to emerging processor technologies like multicore CPUs, GPUs, manycore processors, FPGAs and recent coprocessor boards. It integrates modern numerical concepts and vendor-specific and cross-platform programming environments for best performance, optimal resource utilization and productivity for application developers assuming an environmement based on high performance computing.

Hardware-aware numerics is a multi-disciplinary approach to identify the best combination of applications, physical models, numerical schemes, parallel algorithms, and platform-specific implementations that is giving the fastest and most accurate results on a particular platform. Hardware-aware numerics not only comprises highly-optimized platform-specific implementations, design of communication-optimized data structures, and maximizing data reuse by temporal and spatial blocking techniques – it also relies on the choice and development of adequate mathematical models that express multilevel-parallelism and scalability up to a huge number of cores.