Performance of Large-Scale Electronic Structure Calculations on Built-in FPGA Systems

We discuss the feasibility of an in-house Schrödinger equation solver on the Intel Broadwell Xeon processor with a built-in FPGA, with a particular focus on the performance of large-scale sparse matrix-vector multiplication (SpMV) that is the core numerical operation of electronic structure simulations for multi-million atomic systems. The double-precision SpMV section in our solver is offloaded to FPGA using OpenCL SDK tool chain presented by Intel corporation. Memory-pinning and duplication of compute unit are employed to improve the performance, where we find that memory-pinning and duplication of compute units lead 1.64x faster data-transfer and 1.58x speed-up of SpMV, respectively, against the result with a single compute unit and no pinned memory.