BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CECS - ECPv6.18.4//NONSGML v1.0//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-ORIGINAL-URL:https://www.cecs.uci.edu
X-WR-CALDESC:Events for CECS
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-Robots-Tag:noindex
X-PUBLISHED-TTL:PT1H
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:20220313T100000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:20221106T090000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:20230312T100000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:20231105T090000
END:STANDARD
BEGIN:DAYLIGHT
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
TZNAME:PDT
DTSTART:20240310T100000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
TZNAME:PST
DTSTART:20241103T090000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTART;TZID=America/Los_Angeles:20231103T103000
DTEND;TZID=America/Los_Angeles:20231103T113000
DTSTAMP:20231101T164120Z
CREATED:20231031T234255Z
LAST-MODIFIED:20231101T164120Z
UID:15894-1699007400-1699011000@www.cecs.uci.edu
SUMMARY:Sparse Computing and Large Language Model: from AI 1.0 to AI 2.0
DESCRIPTION:Title: Sparse Computing and Large Language Model: from AI 1.0 to AI 2.0 \nSpeaker: Prof. Guohao Dai \nDate and Time: Friday\, November 3rd\, 10:30 am \nLocation: ISEB 4020 \nAbstract: \nGeneral Artificial Intelligence is going through a new development stage from Computation AI\, Perception AI\, to Cognition AI\, with the new trends of big data\, multi-modality\, and multi-task. However\, due to the slowdown of Moore's Law\, there are bottlenecks in hardware computility\, leaving a huge gap between the supply of computility and the demand for the computility in intelligent computing. Because the AI function has the learnable characteristics\, the amount of calculation can be reduced through the sparsification method\, thereby breaking through the hardware computility bottleneck of intelligent computing while ensuring task accuracy. This report will focus on the software and hardware co-design method of sparse computing. This report will introduce the definition and process of sparse computing\, as well as the challenges faced in intelligent computing from algorithm\, compilation\, and architecture perspectives. This report will elaborate on the multi-level design method for sparse computing in AI\, and introduce three typical models in the AI ​​1.0 era (Graph Neural Networks\, Point Cloud Networks\, and Conventional Convolution Neural Networks). Finally\, this report takes the Large Language Models (LLMs) algorithm in the AI ​​2.0 era as an example to illustrate the application of the above design method in large language model inference optimization. Compared with the existing model\, the calculation speed on dense hardware architecture is increased by 2 times. \nBiography: \nGuohao Dai is a tenure-track Associate Professor at Shanghai Jiao Tong University\, he received his B.S. and Ph.D. (with honor) degrees from Tsinghua University\, Beijing\, in 2014 and 2019. Prof. Guohao Dai has published more than 50\nhigh-level international journals and conference papers in the fields of Electronic Design Automation (EDA)\, heterogeneous computing\, and system/architecture design. He has received more than 1000 citations in Google Scholar. The published paper won the ASP-DAC 2019 Best Paper Award\, DATE 2023/DAC 2022/DATE 2018 Best Paper Nomination\, and WAIC 2022 Outstanding Youth Paper Award. He has personally won honors such as the WAIC 2022 Yunfan Award\, the global champion of the NeurIPS21 BIGANN competition\, Beijing’s outstanding doctoral graduates\, Tsinghua University's outstanding doctoral graduates\, and Tsinghua University’s outstanding doctoral thesis. Professor Dai has participated in guiding students to rank third in the world in ACM 2021 SRC and first in the world in MICRO 2020 SRC. Currently\, he serves as PI/Co-PI for several projects with a personal share of over RMB 10 million. \nHosted By: Sitao Huang \nMore Information
URL:https://www.cecs.uci.edu/event/sparse-computing-and-large-language-model-from-ai-1-0-to-ai-2-0/
LOCATION:ISEB 4020\, Interdisciplinary Science and Engineering Building\, Irvine\, CA\, 92697\, United States
CATEGORIES:Seminars at CECS
ATTACH;FMTTYPE=image/jpeg:https://cecs.uci.edu/files/2023/10/Guohao_Dai_smaller_11-3-23.jpg
END:VEVENT
BEGIN:VEVENT
DTSTART;TZID=America/Los_Angeles:20231130T093000
DTEND;TZID=America/Los_Angeles:20231130T103000
DTSTAMP:20240104T190401Z
CREATED:20240104T185923Z
LAST-MODIFIED:20240104T190401Z
UID:15937-1701336600-1701340200@www.cecs.uci.edu
SUMMARY:Towards Finding the Optimization Potential of Fine-Grained Locks
DESCRIPTION:Name:  Ahmad Nahian \nChair: Professor Brian Demsky \nDate: November 30\, 2023 \nTime:  9:30 AM \nLocation: ISEB 1010 \nCommittee: Professor Rainer Doemer\, Professor Quoc-Viet Dang \nTitle: Towards Finding the Optimization Potential of Fine-Grained Locks \nAbstract: Fine-grained locking is an effective way of eliminating lock contention in multi-threaded applications. However\, it is comparatively difficult and error prone to implement. To help programmers know the optimization potential of implementing fine-grained locking we have developed an approach that overcomes several shortcomings of prior approaches like record/replay. Our approach executes the existing implementation of the multi-threaded program with a custom thread scheduler that schedules threads according to their logical time\, which is the time it would have taken to execute the thread without waiting for unnecessary lock contention. The custom thread scheduler leverages a prediction mechanism to predict conflict relations among critical sections so that it schedules threads to generate the execution of an implementation with optimal fine-grained locking. We have implemented our approach in FGLPerf and tested it on several multi-threaded benchmarks. \nShort Bio: Ahmad Nahian is a PhD candidate in the EECS Computer Engineering program at UC Irvine\, working under the supervision of Professor Brian Demsky. Ahmad Nahian attained his BS in Computer Science and Engineering from Bangladesh University of Engineering and Technology and MS in Computer Engineering from UC Irvine. A notable work of his research is information flow profiler that finds optimization opportunities in multi-threaded programs. The information flow profiler is a powerful tool in that it can find new optimization opportunities along with optimization insights to eliminate synchronization bottlenecks that cannot be discovered by other state of the art profilers.
URL:https://www.cecs.uci.edu/event/towards-finding-the-optimization-potential-of-fine-grained-locks/
LOCATION:ISEB 1010\, Interdisciplinary Science and Engineering Building\, University of California\, Irvie\, CA\, 92697\, United States
CATEGORIES:PhD Defenses
END:VEVENT
END:VCALENDAR