A CUDA-Based Implementation of a Fast Response Urban Wind Model

A new dispersion modeling system based on the well-used FORTRAN-based QUIC (Quick Urban and Industrial Complex) dispersion modeling system has been under development to provide high-resolution wind and concentration fields in cities. The fast response 3D diagnostic urban wind model written in C++, CUDA-URB, is a rapid mass conserving wind-field solver. CUDA-URB uses a variational analysis technique to ensure the conservation of mass rather than slower yet more physics-based solvers that include conservation of momentum. CUDA-URB minimizes the difference between an initial wind field that is specified using empirical parameterizations and the final wind field. This method requires the solution of a Poisson equation for Lagrange multipliers. The Poisson equation is solved using the Successive Over-Relaxation (SOR) method (an iterative solver), which is a variant of the Gauss-Seidel method with more rapid convergence. CUDA-URB utilizes the concept of dynamic parallelism in NVIDIA’s parallel computing-based Graphics Processing Unit (or GPU) API, CUDA, to substantially accelerate wind simulations. CUDA-URB’s GPU-based solver is 25-80 times faster than the equivalent CPU (Computing Processing Unit, not parallelized) solver for domains with total number of 500000-90000000 cells. This enables CUDA-URB to handle larger domains in much less time. As a result, CUDA-URB is able to solve for higher resolution winds in urban areas.

Another feature that makes CUDA-URB an appropriate choice for modeling high-resolution (~ 1-3 m) winds in urban areas, is the addition of a new way of meshing computational domains. We use the cut-cell method, which provides the capability of having cells partially cut by solid surfaces. This method is able to capture more geometric details as compared to standard QUIC-URB’s “block buildings” or other stair-step approaches. The cut-cell method is also applied to topography, affording a more accurate representation of urban terrain and buildings. Results using CUDA-URB’s cut-cell method have been validated against experimental results for terrain and building cases. The results show more accurate wind fields in urban environments.