Batmobile: 10-20x Faster CUDA Kernels for Equivariant Graph Neural Networks (elliotarledge.com)

The recently announced "Batmobile" project introduces custom CUDA kernels that significantly enhance the performance of equivariant graph neural networks (GNNs), achieving speed increases of 10-20x for critical operations like spherical harmonics and tensor products. Equivariant GNNs, while being mathematically elegant and effective for complex tasks in molecular dynamics and drug discovery, have historically been hampered by computational inefficiencies, with operations taking considerable time during model inference. Batmobile addresses these bottlenecks through specialized CUDA kernels that utilize compile-time constants, compute directly in GPU registers, and fuse operations, resulting in dramatic improvements in execution speed. The implications for the AI/ML community are profound, as this optimization makes it feasible to apply equivariant GNNs to real-world applications that involve processing massive datasets, such as molecular simulations and material discoveries. For instance, Batmobile's benchmarks demonstrate that it can compute spherical harmonics and tensor products at speeds 11.8x and 20.8x faster, respectively, compared to the existing e3nn library. This advancement not only enhances the usability of these sophisticated models but also broadens the scope of problems they can tackle efficiently, marking a significant step forward in the computational resources required for atomistic machine learning.