Explorando o Potencial e Limitações do ChatGPT para Paralelização de Códigos: Uma Análise Empírica de Desempenho e Eficiência

Abstract

Parallelization is a well-known strategy for optimizing algorithms, yet its effective imple mentation is fraught with challenges. In this context, ChatGPT, a language model trained on vast datasets, is investigated as a potential tool for high-performance programming. This work is dedicated to exploring the capabilities and limitations of ChatGPT in code parallelization, a critical area for enhancing computational performance in the face of in creasing demands for efficiency. Through a case study, practical experiments, and quanti tative evaluations, we aim not only to highlight the benefits and challenges of integrating ChatGPT into parallelization but also to establish a methodology for future research in the f ield. The experiments were conducted focusing on two problems. For each problem, com parative tests between sequential and parallel implementations were carried out, assessing the performance and efficacy of the solutions generated by ChatGPT. The results indicate that while ChatGPT can generate efficient and optimized parallel codes, the quality of pa rallelization significantly depends on the complexity of the problem and the instructions provided to the model. Where instructions were clear and specific, an improvement in performance was observed. Conversely, vague or generic instructions resulted in less op timized solutions. We conclude that ChatGPT holds significant potential as a tool to aid in code parallelization, but it is imperative that users provide precise and detailed instruc tions and have prior knowledge of their objectives to fully leverage its capabilities. This study serves as a starting point for future investigations, aiming to enhance the integration between artificial intelligence and high-performance programming.