Abstract

In-Database ML – integrating ML/AI features directly into DBMS –becomes more and more a requirement of customers and system builders. This includes system tasks such as indexing and query optimization, as well as application support such as inference and advanced data analytics. Although GPUs are an established technology for accelerating machine learning (ML) tasks, they are costly and require data transfer between the host and device. In contrast, neural processing units (NPUs) are becoming increasingly attractive, even as CPU on-chip extensions. In this work, we study if and how NPUs – which are dedicated to neural network operations – can be leveraged for database operations. We discuss opportunities and present results of our experiments for selected database use cases. Our findings indicate that NPUs are well-suited for latencysensitive, small-scale ML tasks in database systems, such as learned indexes or ML-based operators.