Challenges in GPU-Accelerated Nonlinear Dynamic Analysis for Structural Systems

Many numerical simulation methods, such as finite-element analysis, were originally formulated to run serially or in parallel on central processing units (CPUs). However, computer engineering has seen a paradigm shift toward massive parallelism using graphics processing units (GPUs), which have become the default accelerators in many data-driven scientific disciplines outside of civil engineering. This state-of-the-art review highlights the challenges and practicalities of GPU-accelerating nonlinear dynamic analyses for civil structural problems. To demonstrate the feasibility of a fully GPU-accelerated finite-element analysis, a GPU-based program for linear-elastic dynamic analysis was implemented, where all stages of the analysis were ported to the GPU. Observed speedups were 115 times that of an equivalent CPU-driven analysis for 106 model degrees of freedom (dof). Importantly, the computational time for the assembly and update levels of the analyses were nearly independent of the number of dof. High-resolution simulations of complex structures can be computationally expensive, but these results and advances in other fields suggest that some levels of the finite-element analysis of civil structures could be accelerated using GPUs at increased model resolution with little increase in computational cost, demonstrating the potential for GPU-accelerated computing. However, compared to other GPU-accelerated finite-element analysis applications, the dynamic analysis of civil structures is subject to unique challenges that need to be addressed before GPU acceleration can be fully realized. Aspects of simulating the response of civil structures considering nonlinear response under extreme loading may not be immediately amenable to GPU acceleration; e.g., the use of many differing element formulations within a model, potential for inelastic response and varying degrees of nonlinearity across elements, and traditional reliance on implicit integration schemes with direct solvers. The shift to GPUs is part of a larger movement toward specialized hardware using fine-grained parallelism, and structural engineers need to address these challenges as these emerging technologies become more prevalent.