Since early May, we’ve had a brand-new simulation computer equipped with an AMD Threadripper and an Nvidia RTX3090 at Dahlem Beratende Ingenieure! Naturally, one of our first actions was running a simulation with SplashTool_GPU as a performance test.
First impressions: The GPU is incredibly fast! Even large areas exceeding 100 km² can be computed at a 1-meter resolution in just a few hours. What’s happening under the hood can be nicely visualized using the monitoring tool GPU-Z:
During iteration, the GPU is utilized at 97%. Thanks to the optimization of the CUDA kernel, memory speed is no longer a bottleneck, allowing nearly 100% of the GPU to be utilized while the memory controller still has some buffer.
Additional details can be gleaned from the graph: The brief drop in GPU utilization is caused by the evaluation of the Iteration_Checkpoint. This evaluation runs on the CPU rather than the GPU to minimize GPU memory usage. The drop is very short, and its frequency can be manually adjusted by the user by modifying the iteration steps until the next checkpoint. The longer drop results from writing output files. This takes a few seconds since GeoTif files are written to the hard drive serially and haven’t been parallelized yet. Because non-parallelized output uses less RAM, is simpler to implement, and output files aren’t typically written often, this method remains unchanged.
During simulation, the GPU draws a solid 300 watts, and the system can easily dissipate the heat with fans running at less than 50% speed. This highlights how crucial it is to invest not just in the graphics card, but in the entire system. The case and extra fans are designed to maximize heat dissipation, and the computer sits in a climate-controlled server room. With this setup, continuous operation poses no issues.