# 面壁智能开源 ForgeStencil 自动优化 HPC 应用

- 来源：OpenBMB (@OpenBMB)
- 发布时间：2026-08-04 21:38
- AIHOT 分数：59
- AIHOT 链接：https://aihot.virxact.com/items/cmseqdidu137vro2ecqnk42zo
- 原文链接：https://x.com/OpenBMB/status/2084635010934677684

## AI 摘要

面壁智能（OpenBMB）发布开源 AI 系统 ForgeStencil，无需人工干预即可自动研究并部署 Stencil 优化，一周内在 100+ 工业和科学计算应用中实现中位数 1.41×、几何平均 2.05× 的端到端加速。系统由 Kernel Agent 和 App Agent 两个 AI 智能体驱动，已应用于天气预报、地震成像、电磁仿真、核工程、医学影像和量化金融等真实负载。

## 正文

🚀 AI is moving beyond code generation.

We're releasing ForgeStencil - the first open-source AI system designed to automatically research and deploy Stencil optimizations for real-world HPC applications. 🥳

With no human in the loop， ForgeStencil autonomously researches， optimizes， and deploys stencil kernels - delivering a median 1.41× / geometric-mean 2.05× end-to-end speedup across 100+ industrial and scientific computing applications in just one week.

Powered by two AI agents， ForgeStencil automates the full optimization workflow：
🧠 Kernel Agent researches optimization strategies and generates high-performance kernels.
⚙️ App Agent identifies performance hotspots， integrates optimized kernels into applications， and validates correctness.

ForgeStencil has been applied to real-world workloads including：
🌦️ Weather forecasting
🛢️ Seismic imaging
📡 Electromagnetic simulation
⚛️ Nuclear engineering
🏥 Medical imaging
💹 Quantitative finance

By automating Stencil optimization for real industrial and scientific applications， ForgeStencil enables scalable optimization beyond traditional expert-driven workflows.

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

⭐ Star the repo， try it out， and share what you'd like to optimize with ForgeStencil.