Lattice QCD as a key benchmark for exascale systems

In this talk, I discuss benchmarks of the Wilson-Dirac operator in lattice QCD on large-scale GPU systems from the point of view of the roofline model. Lattice QCD benchmarks have a medium arithmetic intensity and are typically bound by the network bandwidth on contemporary systems. One important implication is that, to first approximation, benchmark results become independent of the GPU model at large scale, which is supported by benchmark data. Traditional Top500 benchmarks of HPC systems are performed using HPL (very high arithmetic intensity) and HPCG (very low arithmetic intensity). Despite benchmarking sparse matrix multiplications, HPCG poorly constrains network bandwidth and is mainly bound by memory bandwidth, making lattice QCD benchmarks a strong candidate for ranking the fabric of a system.