High-Performance Moment-Encoded Lattice Boltzmann Method with Stability-Guided Quantization

TL;DR

This work addresses high-memory, high-computation constraints in lattice Boltzmann simulations with fluid–solid coupling by introducing a split-kernel GPU design and a stability-guided 16-bit moment quantization for HOME-LBM. By decoupling fluid evolution from solid interactions and performing von Neumann stability analysis tailored to HOME-LBM, the authors derive moment-wise stability bounds that enable robust fixed-point quantization and substantial memory reductions without sacrificing fidelity. The approach yields up to several-fold speedups and strong scalability on a single GPU, demonstrated on large-scale fluid-only and fluid–solid benchmarks, including real-time demonstrations with complex geometries. The results highlight the practical impact of combining theoretical stability insights with GPU-oriented algorithm redesign to support high-resolution, geometry-rich, and real-time LBM simulations.

Abstract

In this work, we present a memory-efficient, high-performance GPU framework for moment-based lattice Boltzmann methods (LBM) with fluid-solid coupling. We introduce a split-kernel scheme that decouples fluid updates from solid boundary handling, substantially reducing warp divergence and improving utilization on GPUs. We further perform the first von Neumann stability analysis of the high-order moment-encoded LBM (HOME-LBM) formulation, characterizing its spectral behavior and deriving stability bounds for individual moment components. These theoretical insights directly guide a practical 16-bit moment quantization without compromising numerical stability. Our framework achieves up to 6x speedup and reduces GPU memory footprint by up to 50% in fluid-only scenarios and 25% in scenes with complex solid boundaries compared to the state-of-the-art LBM solver, while preserving physical fidelity across a range of large-scale benchmarks and real-time demonstrations. The proposed approach enables scalable, stable, and high-resolution LBM simulation on a single GPU, bridging theoretical stability analysis with practical GPU algorithm design.

Figures (13)

• Figure 1: 3D Smoke over an F1 Car. Smoke around an F1 car highlights a complex, multi-scale wake: strong shear layers from the wheels, rear wing, and diffuser roll up into coherent vortices near the vehicle, then stretch and break down into a turbulent, filamentary plume downstream.
• Figure 2: 3D High-Resolution Plume. Fluid-only plume simulation using 16-bit quantized kinetic solver at a grid resolution of $1024^3$. Left: Reynolds number $Re = 40,960$. Right: Reynolds number $Re = 2,048,000$.
• Figure 3: 3D Hilbert Space-Filling Curve. Visualization of four different Hilbert curve levels mapped onto a volumetric domain with a grid resolution of $256 \times 512 \times 256$. From left to right, the images correspond to increasing curve refinement levels, illustrating progressively finer spatial traversal patterns within the volume.
• Figure 4: Solid boundary intersection. Illustration of fluid lattice nodes intersecting with a solid surface.
• Figure 5: 3D Yeahright. Comparison of flow simulations using different lattice velocity models and mesh resolutions, including D3Q27 with a low-resolution mesh, D3Q27 with a high-resolution mesh, and D3Q19 with a high-resolution mesh.