Frequently Asked Questions about the GPU Solver

What is changing behind-the-scenes when I select the GPU solver?

Nearly all of the model pre-processing and post-processing still occurs within the CPU when the GPU solver is selected. The TVD solver scheme will be utilised, and your simulation timesteps should be adjusted accordingly. Please note that the ADI solver (and not the TVD solver) is selected by default for CPU simulations.

Will I get identical results using the GPU solver?

The GPU solver utilises the TVD solver scheme. This solver scheme is also available for CPU-based simulation runs, however the ADI solver scheme will be selected by default. Note that smaller timesteps are required for the TVD solver.

Due to differences in the algorithms of the CPU and GPU-based TVD computations, very small differences may occur between their results. Numerical operations are slightly different (bit-level) on GPU when compared to TVD.

Note that, as with all numerical solutions to real-life problems, even tiny differences in one location may lead to large differences elsewhere.

Can I run integrated simulations using the GPU solver?

Yes, the GPU solver is compatible with integrated simulations, including those linked to the 1D river solver, the 1D urban solver, or both! Please note that the relationship between your timesteps will likely be very different to CPU-based simulation runs (i.e. your 1D timestep may be bigger than your 2D timestep!)

2D models with embedded 1D structures have also been integrated into the GPU solver. The GPU solver does not support links to TUFLOW's 1D solver (ESTRY).

Can I run Flood Modeller's GPU solver in Flood Cloud?

Yes, Flood Modeller’s 2D GPU Solver can be run within Flood Cloud and provides access to a NVIDIA Tesla V100 GPU card.

I have multiple GPU cards - what are my options?

It is possible to select different GPU (by its ID – 0-based count) when running multiple simulations. Users currently need to manually edit the xml file or use Windows Command Prompt to launch the 2D engine. This is not currently possible via the interface. You are unable to use more than one GPU in one simulation.

GPU 2D timestepping

As with the CPU-based TVD solver, the selection of the simulation time step affects the numerical stability of the GPU-based computations. To ensure that the simulation's maximum Courant number runs around 1, the ratio of simulation time step to the domain resolution needs to be relatively small (~1/10).

GPU 1D timestepping

The selection of the time step for a linked 1D flow model (Urban/River) needs to ensure (a) the convergence of the flow solution for the 1D model and (b) the timesteps are multiples of each other with the time step of the 2D model being smaller than that of the 1D model.

Furthermore, because the solvers for the linked 1D models run on CPU, the selection of the 1D time step can affect the run time of the simulation. Selecting 1D time-steps that are multiple times larger than the time step of the linked 2D GPU-resolved model can result in smaller run times.

More detailed information about the selection of the 1D time in relation to numerical stability of 1D River models is available in the Adjusting advanced parameters to improve 1D model stability section of this manual.