ATHENA SEMINAR SERIES: Processing-in-memory Accelerators Toward Energy-Efficient Real-World Machine Learning

Presenter: Bokyung Kim

The NSF AI Institute for Edge Computing (Athena) presents the next in our seminar series by Bokyung Kim, titled “Processing-in-memory Accelerators Toward Energy-Efficient Real-World Machine Learning” on Thursday, October 19, 2023, from 12-1pm EST via Zoomhttps://duke.zoom.us/meeting/register/tJMsduyuqzopGNQOjO_eGp4lD2QwoFJ50mXk and in-person at Duke University.

Abstract:

Unlike the decent advance of machine learning algorithms, hardware development falls far behind because of the separation of storage and computation. Processing-in-memory (PIM) accelerators have appeared to overcome the stagnant advancement in hardware by infusing the processing capability into memories. PIM designers should consider numerous factors, from low levels of devices and circuits to high levels with algorithms and applications, to achieve efficiency and reliability of hardware.

In this talk, I will explain energy-efficient PIM-based architecture, chip, and system designs with diverse memory types to accelerate machine learning algorithms. Specifically, I will start from 1) PIM designs with resistive random-access memory (RRAM)—a representative emerging nonvolatile memory (eNVM)—for deep learning models. Two works provide insight into the importance of the 3D architecture and dataflow in PIM accelerators. However, while RRAM is a promising solution for the next generation, the endurance issue still limits its application to the real world. With conventional memory technologies, PIM designs can be leveraged for specific applications showing higher resource requirements and restricted environments. I will detail 2) a fabricated SRAM-PIM chip for diagnosis automation, which can detect and predict disease with high speed and low power as an implementable device for patients. Lastly, I will suggest a direction for PIM designs in our machine-learning ubiquitous lives—3) DRAM-PIM architecture for privacy.

 

Bio:

Bokyung Kim is a Ph.D. candidate in Electrical and Computer Engineering at Duke University under the supervision of Dr. Hai (Helen) Li. She graduated with honors from Ewha Womans University in South Korea, where she received her M.S. and B.S. degrees in Electrical and Computer Engineering and Electronic and Electrical Engineering, respectively. Her research area focuses on efficient processing-in-memory accelerators for machine learning models. She has broad experience in hardware design, spanning different system levels through device modeling, mixed-signal VLSI design, and chip fabrication. She won an NSF iREDEFINE professional development award from the ECE Department Heads Association and is a select fellow of EECS Rising Stars. She will earn her Ph.D. degree in the Spring of 2024.