On August 4, ModelBest and OpenBMB open-sourced ForgeStencil, a two-agent system for stencil optimization. It finds hotspots in real scientific and industrial software, forges CUDA operators, integrates patches back into the application, then measures correctness and speed with the application's own harness.
The official README says ForgeStencil has run on 100 end-to-end validated applications across oil and gas, electromagnetics, medical imaging, CFD, climate, astrophysics, materials and HPC benchmarks. The reported median end-to-end speedup is 1.41x, with a 2.05x geometric mean.
The hard part is deployment
Stencil code sits under weather simulation, seismic imaging, electromagnetic modeling and fluid dynamics. It repeatedly reads neighboring grid points and updates the current point, so poor memory layout can leave a GPU waiting on data.
ForgeStencil splits the work between a Kernel Agent and an App Agent. The first explores tiling, fusion, layout and occupancy strategies. The second locates application hotspots, builds a baseline, verifies correctness and integrates the result back into the real program.
“ForgeStencil automates the strategy discovery and the deployment.”
Read the numbers by scope
For the 100 application results, the README reports a range from 0.998x to 97.7x. It says 89% of apps exceed 1.05x speedup, 73% exceed 1.20x, 43% exceed 1.50x, 21% exceed 3x and 7% exceed 10x.
Operator-level results are separate. On A100, ForgeStencil's memory-bound kernels reach 74% to 85% of peak DRAM bandwidth, and its same-precision comparison across 32 cases shows about 2.16x geometric-mean advantage over the best public baselines.
Chinese coverage also points to concrete applications: gprMax/FDTD Maxwell at 2.47x, minisweep at 5.78x, RTM seismic imaging at 1.81x, non-Cartesian MRI at 2.45x and DBT backprojection at 1.63x.
The next test is reproduction
The useful signal is not the highest speedup. It is whether outside teams can reproduce the fresh-clone workflow, add new applications in patch mode, and see end-to-end validation on H100 and B200. The 100 application packages are currently validated on A100, while broader end-to-end validation is still on the roadmap.
For manufacturers and scientific-computing teams, ForgeStencil matters because it turns scarce HPC tuning knowledge into a recorded, auditable agent workflow. That is a harder claim than “AI wrote code,” and a more useful one if the numbers hold up outside the release environment.
Sources: OpenBMB ForgeStencil GitHub README, CocoLoop, Xinzhiyuan; sources verify the 100 validated applications, A100 end-to-end speedups, operator-level baselines, FDTD/minisweep/RTM/MRI/DBT application scopes and validation limits.