SABLE GPU accelerator speeds power flow 253x for AI training

Researchers from the team of Suho Park, Keunju Song, and Hongseok Kim have introduced SABLE, a GPU-powered accelerator for solving AC power flow problems in a batched, differentiable manner. The key innovation is a block-diagonal embedding that reformulates batched three-dimensional Jacobians into a fixed-pattern two-dimensional sparse template. This template is shared across PyTorch, CuPy, and cuDSS, enabling zero-copy interoperability and memory-efficient sparse reuse. By leveraging reusable sparse templates, custom GPU kernels, a cuDSS-based sparse-direct LU solver, and mixed-precision techniques, SABLE dramatically speeds up repeated power flow computations.

In extensive benchmarks, standalone power flow solving throughput improved by up to 253.4× over pandapower and 5.7× over ExaPF. For end-to-end training using AC optimal power flow learning models (DC3 and DeepLDE), SABLE expanded the feasible training batch range by up to 64× and boosted training throughput by up to 206.7× over baselines. This makes SABLE a powerful tool for integrating physics-based power flow layers into modern deep learning frameworks, enabling faster, more scalable grid optimization and AI-driven energy management.