Abstract: Eulerian Fluid Simulation is an effective formula of modeling fluid in many HPC applications. The high computational complexity of traditional solvers leads to limited deployment, while several neural-network-based methods reduce the labor of computation by learning complicated relationships and draw wide attention. However, these state-of-the-arts apply single network topology to this input-sensitive application, resulting in inability for achieving required accuracy or inefficiency for leveraging the power of machine learning.
In this work, we propose Smart-fluidnet, which provides various approximate models and enables adaptive approximation to accelerate fluid simulation. We generate multiple network typologies by designed transformation operations and select the most suitable models by offline output-quality controller, then a quality-aware runtime algorithm dynamically singles out the optimal one with ignorable runtime overhead. The results show that Smart-fluidnet is 46% and 590x faster than Tompson's method (the state-of-the-art neural network model) and original fluid simulation respectively on an NVIDIA Pascal GPU.
Back to Technical Papers Archive Listing