NextFin News — ModelBest and the OpenBMB community open-sourced ForgeStencil on Tuesday, the first AI system to automate both research and deployment of stencil optimizations in real industrial and scientific software.
The system pairs a Kernel Agent that researches and generates optimized CUDA stencil operators with an App Agent that locates hotspots, forges application-specific operators from the resulting knowledge base, verifies correctness and integrates the changes. It has completed end-to-end validation on 100 applications spanning oil and gas, electromagnetics, medical imaging, computational fluid dynamics, climate modeling, astrophysics, materials science and high-performance computing benchmarks. Measured against each application’s own production GPU code, the system delivered a median 1.41× end-to-end speedup and a geometric-mean 2.05× speedup, with individual gains ranging up to nearly 100× and 89 percent of cases exceeding 1.05×. All results are reproducible from a fresh clone under an auditable harness that interleaves timing and applies the program’s native correctness checks.
Stencil computations form the computational core of many simulation and modeling workloads, yet prior tools automated only code generation while leaving strategy discovery and integration to human experts. ForgeStencil removes that remaining manual loop, packaging operators under Apache-2.0 with runtime dispatch across A100, H100 and B200 architectures and shipping only integration patches rather than third-party source. The release extends ModelBest’s earlier ForgeTrain framework and supplies operators and harnesses that industrial and scientific software teams can apply directly, converting weeks of expert tuning into automated cycles measured in days.
Explore more exclusive insights at nextfin.ai.
