A new weather model runs fast on a single graphics processing unit
The WRF-JAX v1.1.28 model matches the performance of 128-process CPU runs using a single GPU.
Researchers created WRF-JAX v1.1.28. This is a differentiable regional weather model. It uses fully coupled physics.
The model is a JAX reimplementation of the WRF 4.5 dynamical core. It includes the complete CONUS physics suite. The team conducted 27 ten-day integrations across nine regions and three initialization months.
The model uses a horizontal grid spacing of 3 km and grid points. A dynamical core is the part of the model that handles movement. Physics refers to the physical processes included in the model.
The mean sustained integration time on a single NVIDIA A800 was 175.0 s per model hour. This matches the speed of 128-process single-precision WRF runs. The model provides computable approximate gradients for regional weather simulation.
The experiments backpropagated through 100 steps with full physics coupling. This allowed them to obtain all three-dimensional gradient components of the initial UVTQ fields in a single calculation.
Key facts
- WRF-JAX v1.1.28 matches the speed of 128-process CPU runs on a single GPU.
- The model uses a horizontal grid spacing of 3 km and grid points.
- The model provides computable approximate gradients for regional weather simulation and end-to-end optimization.