ForgeStencil – ForgeStencil: FaceWall Intelligence’s Fully Automated Stencil Research and Deployment System Operational

ForgeStencil is the world’s first AI optimization system jointly launched by FaceWall Intelligence and the OpenBMB open-source community to support automated research and deployment for Stencil applications. Based on a dual-agent closed loop consisting of the Kernel Agent and App Agent, it enables the entire process—from discovering optimization strategies to integrating with real production software—with zero human intervention. Within one week, the system completed end-to-end optimization for more than 100 industrial and scientific computing software applications, achieving a median speedup of 1.41× and covering 8 major industrial domains and 5 major scientific domains, including oil and gas exploration, electromagnetic simulation, and medical imaging.

Dual-agent closed-loop optimization: The Kernel Agent autonomously researches and synthesizes high-performance CUDA Kernels, while the App Agent handles hotspot localization, operator forging, correctness verification, and application integration.
Automatic strategy discovery: Agents autonomously explore optimization strategies such as tiling, fusion, layout, occupancy, and host-side refactoring, rather than searching within a fixed space.
Deployment in real applications: The system directly optimizes production-grade software and performs end-to-end validation against the application’s own GPU code as the baseline, rather than using simplified benchmarks.
Auditable measurement protocol: A single environment switch alternates between the original and optimized paths. The application’s built-in checks and timers are used to ensure that reported speedups are genuine and reliable.
Cross-generation GPU adaptation: Supports runtime architecture dispatch for A100/H100/B200, allowing the same codebase to automatically select the optimal Kernel path for each architecture.

Follow WeChat and reply “开源” to join the AI open-source project discussion group.
Dual-agent collaborative architecture: The system consists of the Kernel Agent and App Agent, forming a closed loop. The Kernel Agent is responsible for theoretical discovery, autonomously researching and synthesizing Stencil operators that approach the physical limits of the hardware through a Plan→Code→Profile loop. The App Agent handles engineering implementation by locating application performance hotspots, establishing GPU baselines, and using the operator-matrix knowledge base built by the Kernel Agent to forge application-specific optimizations, complete correctness verification, and integrate them into the original application.
Operator matrix and knowledge-base evolution: During optimization, the Kernel Agent continuously builds specialized operator matrices as a knowledge base that subsequent App Agents can reuse. Multiple parallel Agents share this knowledge base, enabling real-time experience sharing and collective synchronized evolution, overcoming the difficulty of scaling the transfer of human experts’ experience.
Auditable end-to-end protocol: Measurement uses the application’s own production-grade GPU implementation as the sole baseline. The original and optimized paths are distinguished only by a single USE_OURS switch, and the original path is byte-for-byte identical to the upstream version. The program’s built-in correctness checks and timers are used, the median is taken after multiple interleaved runs, and the final report gives the geometric-mean speedup across all standard test cases, eliminating opportunities for falsification at the system level.
Patch mode and zero source-code bundling: No third-party source code is bundled. Each application provides only upstream provenance records, a retrieval script, and integration patches. vendor.sh automatically fetches the specified upstream code version, ensuring license compliance and reproducible results.

Environment setup: Clone the repository and ensure that an NVIDIA GPU (validated on A100), CUDA 12.x, a C++17 compiler, and Python 3.9+ are available.
Quickly try an operator: Run python tools/run.py --stencil star_1 --shape 256 --gpu 0 to measure the performance of a single operator within 1 minute and compare it with the Halide baseline.
End-to-end reproduction: Enter the target application directory and run ./vendor.sh to fetch the upstream source code. After applying the application integration patch, run python ../../harness/run_e2e.py --app --gpu auto to complete validation on the real application.
Drive the Agents: Refer to agents/README.md to configure and launch the autonomous optimization loops for the Kernel Agent and App Agent.
Review the results registry: View audited speedup data for 100 verified applications in results/integration_registry.json.

Zero human intervention: The entire process—from application analysis, hotspot localization, and Kernel forging to integration validation—is automated, with no HPC expert decision-making required.
Real-world focus: Directly optimizes production-grade industrial software; 42% consists of real industrial production software, with the application’s own GPU code serving as the baseline.
Dramatically improved R&D efficiency: Reduces optimization time per application from several months to within one day, completing 100+ applications in one week and effectively saving approximately 20–30 person-years of R&D effort.
Strong performance: Achieves a 1.41× end-to-end median speedup; 43% of applications achieve ≥1.5× speedup; and at the operator level, achieves a 2.16× geometric-mean speedup over the best open-source baseline.
Sustainable across architectures: Supports runtime distribution across multiple generations of NVIDIA GPUs and can automatically reforge Kernels during architecture upgrades to capture new hardware capabilities.

GitHub repository: https://github.com/OpenBMB/ForgeStencil

Oil and gas seismic exploration: Optimizes core algorithms such as RTM reverse-time migration and the Total mini-app, accelerating petroleum exploration data processing.
Electromagnetic simulation and design: Improves the efficiency of solving Maxwell’s equations on Yee grids in applications such as gprMax/FDTD, supporting radar and chip electromagnetic analysis.
Medical image reconstruction: Accelerates medical imaging workloads such as non-Cartesian MRI reconstruction and digital breast tomosynthesis backprojection.
Climate and astrophysics: Optimizes Stencil-intensive scientific computing applications, including atmospheric dynamics and cosmological simulations.
Quantitative finance computing: Provides automatic optimization of Stencil hotspots in financial pricing libraries such as QuantLib, reducing the latency of trading computations.