8 engineers spent a year and nearly 10 million yuan in work, and this open source AI system was completed in 7 days

Zhidongzhi Author | Cheng Qian Editor | Yunpeng Zhidongzhi reported on August 4 that today, Face Wall Intelligence and the open source community OpenBMB open sourced the world's first AI optimization system ForgeStencil that supports Stencil (template calculation) automatic research and automatic deployment. According to official disclosures, ForgeStencil completed the automatic optimization of more than 100 real industrial and scientific computing software within a week, which is equivalent to investing about 4-8 engineers per year and equivalent R&D investment of 3 million to 10 million yuan. Generally speaking, experts can fine-tune no more than 20 applications per year. ForgeStencil compresses the optimization time required for a single application to the hour level, improves R&D throughput by about 1-2 orders of magnitude, and increases R&D efficiency by about 100 times. This efficiency can be amplified in parallel as the scale of computing power increases. Breakthroughs in modern industry and cutting-edge scientific research are highly dependent on high-performance computing (HPC) software. However, there is a computing-intensive core computing model Stencil at the bottom of such software. The core of it is to perform the same local operation on each point on a regular grid and its fixed neighborhood. Stencil is a memory-intensive calculation, and its performance is limited by memory bandwidth. It usually takes up most of the application's time. Its optimization level also directly determines the operating efficiency of industrial and scientific research simulation software. Previously, Stencil optimization relied heavily on scarce HPC experts, which also made it difficult to scale. ForgeStencil was born to solve this problem. ForgeStencil is based on Forge Engineering, a new software development idea proposed by wall-facing intelligence. It is another technological breakthrough after its AI-made AI achievement ForgeTrain in May. Open source address: https://github.com/OpenBMB/ForgeStencil 1. The two agents collaborate for tuning without manual intervention. ForgeStencil realizes the optimization from