Optimas: AI-Powered Fully Automated Code Performance Optimization Framework

Optimas is a modular, fully automated end-to-end generative AI framework. It closely integrates performance diagnosis and code transformation via a multi-agent workflow, addressing the pain points of traditional code optimization that relies on experts and has low efficiency. In 3410 real experiments, it achieved 100% code correctness, a 98.82% performance improvement rate, and an average speedup of 8.02% to 79.09%.

In the fields of HPC and GPU acceleration, traditional optimization relies on performance analysis tools to identify bottlenecks, followed by manual optimization by experts—this is time-consuming and experience-dependent. While LLMs bring hope for automation, they lack performance context, making it difficult to generate both correct and effective code transformations.

Optimas' core concept is diagnosis-guided generative optimization, using multi-agent collaboration: the Diagnosis Parsing Agent processes performance reports, the Knowledge Mapping Agent matches optimization techniques, the Code Generation Agent produces targeted modifications, and the Verification Agent compiles and runs to verify correctness and performance.

The experiment scale included 3410 real tests (10 benchmarks, 2 HPC applications, NVIDIA GPU). The results include 100% code correctness, a 98.82% performance improvement rate, and an average speedup of 8.02% to 79.09%.

Technical highlights include multi-report fusion (integrating multi-source diagnostic information), literature-supported optimization library (based on academically validated techniques), automatic verification loop (generation-verification-feedback), and modular design (flexible expansion of tools/techniques).

Applicable users include HPC developers, GPU-accelerated application developers, performance engineers (as a rapid prototyping tool), and CI/CD pipelines (for automatic performance regression and optimization suggestions).

Limitations: Currently mainly supports NVIDIA GPUs, focuses on local code optimization, and requires domain knowledge injection for specific scenarios. Future directions: Expand hardware platforms, introduce algorithm-level optimization, enhance domain adaptability, and explore reinforcement learning applications.

Optimas is expected to democratize performance optimization capabilities, allowing ordinary developers to access high-performance optimization. It has positive impacts on cloud computing (improved resource utilization), scientific research (faster simulation speeds), and AI training (reduced energy consumption), representing a significant breakthrough of generative AI in software engineering.