GPU Accelerated Lattice Boltzmann Method on Apple Silicon via Metal Shading Language: Performance and Energy Efficiency on Unified Memory Architecture

This study natively implements D2Q9 Lattice Boltzmann Method solver on Apple M1 using Metal Shading Language, and quantitatively evaluates the performance and energy efficiency advantages that Unified Memory Architecture (UMA) provides for memory bandwidth bound algorithms. Through both Poiseuille flow and lid driven cavity validation, the present implementation was confirmed to be a physically valid solver that quantitatively agrees with established benchmark data for lid driven cavity flow. In the DRAM bound regime at grid sizes of 512 by 512 and 1024 by 1024, memory bandwidth efficiencies of 86.2% and 87.9% of the theoretical peak were achieved, respectively, which is comparable to existing CUDA and OpenCL based implementations. The central contribution of this study is the quantification of the advantage of UMA, isolated from other implementation factors, through an ablation comparison between a shared memory method and an explicit copy method. The methodinvolving explicit copy operations was consistently 10.2% to 20.3% slower across every grid size tested, and this resultdemonstrates, based on measured data, that the elimination of host device copies by UMA constitutes a structural performance advantage that is not limited to a particular algorithm.