Breaking the Memory Barrier: Accelerating DLRM Inference with Real-World PIM Architecture

Speaker: Dr. ZHOU, Amelie Chi
         Assistant Professor
         Department of Computer Science
         Hong Kong Baptist University

Title:  "Breaking the Memory Barrier: Accelerating DLRM Inference with
        Real-World PIM Architecture"

Date:   Monday, 6 November 2023

Time:   4:00pm - 5:00pm

Venue:  Lecture Theater F
        (Leung Yat Sing Lecture Theater)
        near lift 25/26, HKUST


Abstract:

Deep Learning Recommendation Models (DLRMs) have gained significant
popularity in click-through rate (CTR) prediction and news rankings
applications. However, optimizing the performance of DLRMs is challenging
due to their intensive memory capacity and bandwidth requirements.
Existing studies mainly adopt CPU-GPU hybrid architecture, where large
scale of embedding tables are stored in CPU memories. Due to the high
memory access latencies, various caching techniques are employed to reduce
costly memory accesses. In this talk, we introduce our solution to the
memory bottleneck issue with Process-in-Memory (PIM) accelerated DLRMs.
PIM leverages specialized hardware architectures that integrate processing
units within the memory subsystem. By co-locating computation and data,
PIM architectures can significantly reduce memory access latencies and
improve DLRM inference performance. This talk will delve into the design
principles and implementation details of the PIM accelerated DLRM
approach, highlighting its potential to break the memory barrier and
revolutionize DLRM inference.


*****************
Biography:

Dr. Zhou is an Assistant Professor in Department of Computer Science, Hong
Kong Baptist University. She received her Ph.D. degree in Computer Science
from Nanyang Technological University (NTU) in 2016. She was a postdoc
researcher in INRIA Rennes (2016-2017) and a faculty member of Shenzhen
University (2017-2023). Her research interests include parallel and
distributed systems, cloud computing and high-performance computing. She
has published more than 30 technical articles in refereed journals and
conferences including SC, HPDC, ICS, ICDE, ICDCS, SoCC and TPDS. She has
been actively serving the community by participating in the
organizing/program committees for conferences including SC, HPDC, IPDPS,
Cluster and CIKM. She is also serving as an Associate Editor for IEEE TPDS
and an Editor for FGCS. She is a recipient of the IEEE-CS TCHPC Early
Career Award and the ACM SIGHPC China Rising Star Award in 2021.