FFT, FMM, and multigrid on the road to exascale: Performance challenges and opportunities

Huda Ibeid, Luke N. Olson, William Gropp

Abstract: Highlights•Algorithms that are known to be compute-bound on current architectures, such as the FMM, could become memory-bound on future exascale systems.•Execution time and energy consumption are independent of the algorithmic computational complexity up until the processor balance point. They increase as a function of the algorithmic complexity beyond that point.•It is well known that GPUs deliver more peak performance and bandwidth relative to high-end CPUs. This performance gap is likely to increase towards exascale.•Emerging 3-D stacked DRAM devices will significantly increase available memory bandwidth. However, with the exponential increase in core counts, stacked DRAM will only move the memory wall and is unlikely to break through it.