Introduction

The QiMeng processor chip hardware and software automated design system aims to achieve fully automated design of the chip hardware/software stack by leveraging large language models (LLMs) and agent technologies. QiMeng has achieved automated designing RISC-V CPUs, automated optimizing operating system kernel configuration, automated transcompiling tensor programs and automated developing high-performance libraries, achieving performance that matches or even surpasses that of human expert design.

Creating a new fully automated software and hardware design paradigm for processor chip

Automated processor chip design represents one of the core challenges in computer science. Unlike conventional automated design approaches that use artificial intelligence (AI) as a tool for specific stages of the chip design process, QiMeng system establishes a new fully automated design paradigm for processor chips. Leveraging LLMs and agent technologies, QiMeng aims to achieve fully automated design and optimization for both hardware and software from given functional specifications.

Considering the unique challenges of processor chip design, realizing fully automated hardware and software design requires addressing two critical issues: the correctness guarantee and the enormous solution space. To this end, QiMeng constructs processor chip design agents with two feedback-driven mechanisms: correctness feedback and performance feedback. By constructing automated functional verification and autonomously repairing erroneous results based on correctness feedback, QiMeng is able to ensure the validity of generated outputs. Concurrently, by leveraging automated performance evaluation and autonomously search based on performance feedback, QiMeng is capable of decomposing the solution space and pruning the low-performance subspaces, effectively reducing the dimensionality of the solution space and enabling efficient exploration of high-performance design solutions.

By constructing a domain-specific large processor chip model (LPCM) and further building hardware and software design agents, QiMeng produces automated design methodologies and results for multiple key stages of processor chip hardware/software development.

The automated design methodologies produced by QiMeng matching the performance of human expert design

At present, QiMeng has delivered automated hardware/software design methodologies and results for processor chips, matching or surpassing the performance of human expert design. QiMeng develops the world's first fully automated designed CPU named QiMeng-CPU-v1 (also named Enlightenment-1), and the world's first fully automated designed superscalar processor core named QiMeng-CPU-v2, achieving performance comparable to Intel 486 and ARM Cortex-A53, respectively. QiMeng realizes the world's first LLM-based automated OS kernel configuration optimization method, whose optimized performance surpasses manual tuning by human experts. QiMeng introduces the world's first automated cross-platform tensor program translation tool, enabling automated code translation across different processor architectures and programming models. QiMeng develops the world's first LLM-based automated high-performance matrix multiplication code generation framework and high-performance tensor operator auto-generation framework, the performance of generated operators outperforming that of OpenBLAS and cuBLAS.

QiMeng pioneers novel design methodologies for processor chips. QiMeng holds the potential to surpass manual design and achieve better performance under identical fabrication technology, significantly improving design efficiency while shortening development cycles and lowering costs based on automated design, and enabling rapid customization of chip stacks tailored to specific application domains to address the growing demand for specialized computing solutions.

The Research Achievements Have Been Highly Recognized in the International Academic Community

Multiple related research outcomes of QiMeng have been published at top international conferences and journals in related areas, including Design Automation Conference (DAC) and IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD) which are top conference and journal in integrated circuits, USENIX Symposium on Operating Systems Design and Implementation (OSDI) which is top conference in systems software, International Joint Conference on Artificial Intelligence (IJCAI) and Association for the Advancement of Artificial Intelligence (AAAI) which are top conferences in artificial intelligence, International Conference on Machine Learning (ICML) and Conference on Neural Information Processing Systems (NeurIPS) which are top conferences in machine learning, as well as the Annual Meeting of Association for Computational Linguistics (ACL) and Empirical Methods in Natural Language Processing (EMNLP) which are top conferences in natural language processing. Among these, QiMeng-CPU-v1 (also named Enlightenment-1) was described by Nature News as "good news for Chinese chip development," further demonstrating its exceptional contributions and influence in the field. The domain-specific LLM for automated HDL code generation, CodeV series, has been downloaded over 32K times in open-source communities and has been tracked and cited by leading institutions such as NVIDIA, Intel, IBM, Amazon, Huawei, ETH Zurich, Georgia Tech, UC San Diego, University of Illinois Urbana-Champaign, and the University of Southern California, underscoring its significant impact in the global academic community.