ICSA Colloquium Talk - 08/11/2023

Title: Unlocking the Power of Data-Centric Acceleration for Modern Applications

Abstract: In today's digital landscape, the exponential growth of data has become the driving force behind cutting-edge applications, such as genome analysis and machine learning applications, revolutionizing our approach to healthcare and overall living quality. However, this unprecedented deluge of data poses a formidable challenge to traditional von Neumann computer architectures. The inefficiencies arising from the constant data movement between processors and memory consume a substantial portion of both execution time and energy when running modern applications on conventional von Neumann computers. To reduce this significant data movement, data-centric architectures, particularly processing-in-memory accelerators, emerge as a promising solution by enabling the processing of data directly where it resides. Nonetheless, most existing data-centric architectures primarily focus on accelerating specific arithmetic operations, inadvertently leaving a substantial gap between the architectural enhancements and the holistic needs of modern applications. Concurrently, conventional software optimizations often treat the architecture as a black box, which inherently limits the potential acceleration of applications.

This talk seeks to bridge the gaps between modern applications and data-centric architectures and revolutionize the landscape of data-centric acceleration for two vital categories of modern applications: genome analysis and machine learning. Firstly, this talk offers a comprehensive analysis of the pressing challenges within state-of-the-art genome analysis pipelines and introduces an innovative end-to-end data-centric acceleration approach achieved through seamless software-and-hardware co-design. Secondly, this talk illuminates the path to closing the gap between data-centric accelerators and the execution of real-world applications by presenting a compelling case study centered on a crucial machine learning application based on generative adversarial networks. Furthermore, this talk delves into the intricate challenges of data-centric acceleration for modern applications and explores potential solutions to surmount these obstacles, paving the way for a future where data-centric acceleration seamlessly integrates with the ever-evolving landscape of advanced applications.

Speaker: Dr. Haiyu Mao is a postdoctoral researcher in the SAFARI Research group led by Prof. Onur Mutlu at ETH Zurich, Switzerland. In July 2020, she received her Ph.D. degree in computer science from Tsinghua University, China. Her research interests intersect between computer architecture, processing in memory, bioinformatics, machine learning accelerators, non-volatile memory, and secure memory.  Visit Haiyu’s personal website for more info: https://hybol1993.github.io/.