The Impact of Machine Learning on the Design of Next-Generation Computer Architectures

Abstract

Machine Learning (ML) has typically been considered a software-driven technology functioning on computer systems; however, its growing computational requirements are affecting the design and optimization of computer architectures. This study analyzes the impact of machine learning on the evolution of next-generation computer architectures by reviewing literature related to architecture design supported by machine learning, specialized hardware, and efficient computing resources. Current research shows that ML can be used not only as a task performed by processors but also as a means for architecture design, optimization, forecasting, simulation, and design automation. Studies have also recognized the necessity for tailored architectures that can meet the computational, energy, memory, and latency demands of contemporary ML applications, especially in IoT and edge-computing settings. The literature examined indicates that upcoming computer architectures could depend more on diverse processors, specialized accelerators, co-design of hardware and software, and machine learning-supported design-space exploration. Nonetheless, issues related to computational complexity, energy efficiency, hardware limitations, interpretability, compatibility, and the quality of training data continue to be significant factors. The research finds that machine learning can affect the workloads that computer architectures need to accommodate and the strategies employed to design those architectures.