In various industries, in-database analytics are crucial for decision-making. Yet, the growing amount of data presents challenges as traditional methods become excessively time-consuming and costly. Moore’s Law and Dennard scaling, which previously aided data-intensive analytics, are reaching their physical limits. A new approach is needed to handle analytics workloads.

The speed gap between processing units and memory creates bottlenecks for analytics workloads (the classic von Neumann bottleneck). Our approach involves using “intelligent” memory that can compute results alongside the stored data. While similar methods have been proposed in the past, they have often been one-sided, with either the hardware side proposing a new architecture and hoping the software would adapt, or vice versa. Our work takes a comprehensive approach, collaborating on both hardware and software simultaneously. This talk will present early results from applying this approach to intelligent DRAM and SRAM technologies for analytics workloads. The talk will also introduce an initial DSL that, if implemented in hardware, could unlock the full potential of hardware for analytic workloads across various architectures.

Summer 2023 PDL/SDI Talk Series

Bio:
Jignesh Patel is an incoming professor in the Computer Science Department at Carnegie Mellon University. His research focuses on data management, emphasizing both system efficiency (e.g., making data platforms run faster) and human efficiency (e.g., designing LLM-based query interfaces). His papers have been recognized as the best papers at top database conferences, including SIGMOD and VLDB. He is a fellow of the AAAS, ACM, and IEEE organizations. He has also received teaching awards at the U. of Wisconsin and the U. of Michigan. Jignesh has a strong interest in technology transfer. He has launched four startups and has also made key contributions to product improvements for industry sponsors.

More Info: https://pdl.cmu.edu/talk-series/2023/071323.shtml