Data-flow Conjugate Gradient Solver for Lattice QCD Calculations on FPGA Accelerator

In this talk, we discuss the Lattice QCD Conjugate Gradient solver as data-flow graph. Such a data-flow graph is described in the high-level language MaxJ from Maxeler, which is an openSPL based programming language, to deploy the algorithm on an FPGA accelerator. We show that such an implementation is power efficient and present first power measurement results. In this framework, all operators like the Dslash operator and the spinor field scalar product are deployed as data-flow kernels. Each kernel is a deep arithmetic pipeline and exposes the maximal possible parallelism, thus we reach a high arithmetic intensity. Such a kernel forms a basic block where each block is deployed as piece of hardware and a manager state machine orchestrates the data streams between. In addition, we discuss also the usage of mixed precision number representation like floating point and fixed-point, and present first numerical analysis and convergence tests.