Toshiyuki Imamura

This paper addresses emulation algorithms for matrix multiplication. General Matrix-Matrix Multiplication (GEMM), a fundamental operation in the Basic Linear Algebra Subprograms (BLAS), is typically optimized for specific hardware architectures. The Ozaki scheme is a well-established GEMM-based emulation method for matrix multiplication, wherein input matrices are decomposed into several low-precision components to ensure that the resulting matrix product is computed exactly through numerical operations. This study proposes a novel GEMM-based emulation method for matrix multiplication that leverages the Chinese Remainder Theorem. The proposed method inherits the computational efficiency of highly optimized GEMM routines and further enables control over the number of matrix multiplications, which can enhance computational accuracy. We present numerical experiments featuring INT8 Tensor Core operations on GPUs and FP64 arithmetic on CPUs as case studies. The results demonstrate that FP64 emulation using the proposed method achieves performance levels of up to 7.4 to 9.8 TFLOPS on the NVIDIA RTX 4090 and 56.6 to 80.2 TFLOPS on the NVIDIA GH200, exceeding the measured performance of native FP64 arithmetic. Furthermore, for FP64 computations on CPUs, the proposed method achieved up to a 2.3x speedup in emulating quadruple-precision arithmetic compared to the conventional Ozaki scheme.

We investigate the applicability of the data assimilation (DA) to large eddy simulations based on the lattice Boltzmann method (LBM). We carry out the observing system simulation experiment of a two-dimensional (2D) forced isotropic turbulence, and examine the DA accuracy of the nudging and the local ensemble transform Kalman filter (LETKF) with spatially sparse and noisy observation data of flow fields. The advantage of the LETKF is that it does not require computing spatial interpolation and/or an inverse problem between the macroscopic variables (the density and the pressure) and the velocity distribution function of the LBM, while the nudging introduces additional models for them. The numerical experiments with 256 × 256 grids and 10% observation noise in the velocity showed that the root mean square error of the velocity in the LETKF with 8 × 8 observation points ( ∼ 0.1 % of the total grids) and 64 ensemble members becomes smaller than the observation noise, while the nudging requires an order of magnitude larger number of observation points to achieve the same accuracy. Another advantage of the LETKF is that it well keeps the amplitude of the energy spectrum, while only the phase error becomes larger with more sparse observation. We also see that a lack of observation data in the LETKF produces a spurious energy injection in high wavenumber regimes, leading to numerical instability. Such numerical instability is known as the catastrophic filter divergence problem, which can be suppressed by increasing the number of ensemble members. From these results, it was shown that the LETKF enables robust and accurate DA for the 2D LBM with sparse and noisy observation data.