102 PFLOPS lattice QCD quark solver on Fugaku

Bridge++ 2.0: Benchmark results on supercomputer Fugaku

“…The better performance on Fugaku is due to the higher memory bandwidth (B/F ratio 0.17 for Fugaku, ∼ 0.08 for KNL) and faster MPI reductions. Global reductions on Fugaku with Tofu barrier execute reduction operations up to 3 elements simultaneously with the barrier synchronization, which is in fact efficient in the iterative solvers [14].…”

“…To maximally make use of the hardware potential of the Fugaku supercomputer, so-called co-design development has been adopted that develops hardware and software concurrently against several target applications [13]. Lattice QCD simulation is one of the target applications and the co-design activity accomplished development of the QCD Wide SIMD library (QWS) [14]. The QWS library implements a domain-decompose linear equation solver for the clover fermion and achieved sustained performance of more than 100 PFlops.…”

Wilson matrix kernel for lattice QCD on A64FX architecture

“…The situation just after the start of the shared use of supercomputer Fugaku was reviewed by Y. Nakamura at the symposium on lattice field theory in 2021 (Lattice2021) [10]. The QWS library is a product of co-design project in the development of Fugaku, and achieves sustained performance of more than 100 PFlops [7]. It is a linear equation solver for clover fermion matrix and its implementation uses the Schwarz Alternating Procedure which accelerates the iterative solver by domain-decomposed matrix that reduces communication among the MPI processes.…”

“…It has been also the case for the post-K (supercomputer Fugaku) project in Japan. As one of the target applications of so-called Co-design project [11] in development of the A64FX architecture for Fugaku [5], a high performance lattice QCD library has been developed and released as the "QCD Wide SIMD" (QWS) library [7]. The QWS library implements a linear equation solver, a typical bottleneck in lattice QCD simulations.…”