Home
MegMind Unveils ForgeStencil: Dual-Agent System Automates Optimization of 100+ Industrial Software Tools in One Week
Modern industrial breakthroughs and cutting-edge research rely on large-scale high-performance computing software, with Stencil computation serving as the core computational pattern. This memory-intensive process, constrained by bandwidth, is critical in fields like weather forecasting, seismic exploration, and material science, where speed directly impacts national software efficiency.
Historically, optimizing Stencil patterns required scarce HPC experts and manual tuning, limiting scalability. Mianbi Intelligence and the OpenBMB community have now released ForgeStencil, the world’s first AI-driven Stencil optimization system. This open-source solution automates research and deployment, breaking previous barriers to scalability.

ForgeStencil automatically optimized over 100 industrial and scientific applications in just one week, reducing optimization time to hours and boosting R&D efficiency by 100x. This breakthrough eliminates human constraints, enabling true scalability in Stencil optimization.
Featuring a dual-Agent architecture with zero human intervention, ForgeStencil requires only source code input. The Kernel Agent designs high-performance operators, while the App Agent handles integration and verification. This end-to-end automation ensures correctness without expert involvement.
In operator benchmarks, ForgeStencil outperformed frameworks like Halide and Devito, achieving a 2.35x geometric mean acceleration in fp32 and an additional 1.95x in mixed fp16. It also delivered a 1.34x speedup on complex variable-coefficient Stencil patterns.

ForgeStencil targets real industrial scenarios, accelerating key software such as hypre (3.86x), minisweep (5.78x), and gprMax (2.47x). It directly addresses production needs for oil, medical, and meteorological sectors.
Unlike human experts, ForgeStencil’s parallel Agents share a common knowledge base, enabling collective evolution. Building on the Forge paradigm, it shifts from code generation to automatic optimization and real-world deployment.
Short-term, ForgeStencil reduces costs by optimizing existing software within hours. Long-term, it accelerates digital foundations for CAE, seismic imaging, and chip design. Now open-sourced on GitHub, Mianbi Intelligence invites HPC experts and developers to contribute.
Related article
Inside Claudeforce: Uniting Salesforce Data with Claude AI
Salesforce and Anthropic deepen their strategic alliance with the launch of Claudeforce, a move that enhances enterprise connectivity and platform engagement. Credit: JasonDoiy/Getty ImagesSalesforce and Anthropic expand their strategic partnership a
The sameness problem behind those unappetizing AI-generated menus
At first, you might question your sanity. You step into a café and scan a menu filled with bagel sandwiches, yet every image appears unnervingly perfect—flawlessly symmetrical and overly smooth—triggering an instinctive sense that something is off. Y
IBM Report: AI-Driven Breaches Account for 25% of Incidents, with Individual Losses 20% Above Average
New research from IBM and the Ponemon Institute highlights a disturbing shift: AI-powered attacks now represent 25% of all malicious data breaches, marking a 56% year-over-year surge. These incidents result in an average loss of $6 million, significa
Related Special Topic Recommendations
Comments (0)
0/500
Modern industrial breakthroughs and cutting-edge research rely on large-scale high-performance computing software, with Stencil computation serving as the core computational pattern. This memory-intensive process, constrained by bandwidth, is critical in fields like weather forecasting, seismic exploration, and material science, where speed directly impacts national software efficiency.
Historically, optimizing Stencil patterns required scarce HPC experts and manual tuning, limiting scalability. Mianbi Intelligence and the OpenBMB community have now released ForgeStencil, the world’s first AI-driven Stencil optimization system. This open-source solution automates research and deployment, breaking previous barriers to scalability.

ForgeStencil automatically optimized over 100 industrial and scientific applications in just one week, reducing optimization time to hours and boosting R&D efficiency by 100x. This breakthrough eliminates human constraints, enabling true scalability in Stencil optimization.
Featuring a dual-Agent architecture with zero human intervention, ForgeStencil requires only source code input. The Kernel Agent designs high-performance operators, while the App Agent handles integration and verification. This end-to-end automation ensures correctness without expert involvement.
In operator benchmarks, ForgeStencil outperformed frameworks like Halide and Devito, achieving a 2.35x geometric mean acceleration in fp32 and an additional 1.95x in mixed fp16. It also delivered a 1.34x speedup on complex variable-coefficient Stencil patterns.

ForgeStencil targets real industrial scenarios, accelerating key software such as hypre (3.86x), minisweep (5.78x), and gprMax (2.47x). It directly addresses production needs for oil, medical, and meteorological sectors.
Unlike human experts, ForgeStencil’s parallel Agents share a common knowledge base, enabling collective evolution. Building on the Forge paradigm, it shifts from code generation to automatic optimization and real-world deployment.
Short-term, ForgeStencil reduces costs by optimizing existing software within hours. Long-term, it accelerates digital foundations for CAE, seismic imaging, and chip design. Now open-sourced on GitHub, Mianbi Intelligence invites HPC experts and developers to contribute.
Inside Claudeforce: Uniting Salesforce Data with Claude AI
Salesforce and Anthropic deepen their strategic alliance with the launch of Claudeforce, a move that enhances enterprise connectivity and platform engagement. Credit: JasonDoiy/Getty ImagesSalesforce and Anthropic expand their strategic partnership a
The sameness problem behind those unappetizing AI-generated menus
At first, you might question your sanity. You step into a café and scan a menu filled with bagel sandwiches, yet every image appears unnervingly perfect—flawlessly symmetrical and overly smooth—triggering an instinctive sense that something is off. Y
IBM Report: AI-Driven Breaches Account for 25% of Incidents, with Individual Losses 20% Above Average
New research from IBM and the Ponemon Institute highlights a disturbing shift: AI-powered attacks now represent 25% of all malicious data breaches, marking a 56% year-over-year surge. These incidents result in an average loss of $6 million, significa











