國立虎尾科技大學 |

Composable Architecture Primitives for the Era of Efficient Generalization.

Record Type:	Language materials, manuscript : Monograph/item
Title/Author:	Composable Architecture Primitives for the Era of Efficient Generalization./
Author:	Davies, Michael E.
Description:	1 online resource (143 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
Contained By:	Dissertations Abstracts International85-11B.
Subject:	Computer science. -
Online resource:	click for full text (PQDT)
ISBN:	9798382346885

Composable Architecture Primitives for the Era of Efficient Generalization.
Davies, Michael E.

Composable Architecture Primitives for the Era of Efficient Generalization. - 1 online resource (143 pages)

Source: Dissertations Abstracts International, Volume: 85-11, Section: B.

Thesis (Ph.D.)--The University of Wisconsin - Madison, 2024.

Includes bibliographical references

We are in the age of AI, with rapidly changing algorithms and hardware. As Deep Learning (DL) applications have evolved, architects have turned to specialization to keep up with the insatiable compute that DL demands. Counter to this hardware evolution, DL demands that hardware support a vast number of diverse application behaviors and in addition, an inflexible ossified software interface in order to be viable. This dissertation motivates and adopts the guiding philosophy of "Efficient Generalization" - as opposed to "Specialization" - as a pathway for architects to support diverse DL application behavior while also catering to the ossified DL stack. We take a principled approach to identifying architecture and software primitives to enable performance and efficiency growth while paying careful attention to the restrictions that come with an ossified software stack and reaching the limits of technology scaling. We calibrate our understanding of modern architecture for DL by conducting a longitudinal study of how the DL revolution has panned out for GPUs - the dominant AI hardware platform of the day. Based on the insights from this study, we explore the fundamental architecture primitives needed to support DL's conflicting needs of programmability and high-performance. We use the knowledge of these architecture primitives to then tease apart the diminishing impacts of technology scaling. Combining all our findings about the role of technology, microarchitecture and architecture, we explore a novel execution model, spatial fusion, for DL on existing programmable architectures (GPUs) which can enable better utilization of on-chip resources leading to higher performance and power efficiency with only modest changes to the architecture.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798382346885Subjects--Topical Terms:

573171
Computer science.
Subjects--Index Terms:

Deep LearningIndex Terms--Genre/Form:

554714
Electronic books.

Composable Architecture Primitives for the Era of Efficient Generalization.
LDR:03157ntm a22003977 4500 001 1150156
005 20241022111602.5
006 m o d
007 cr bn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798382346885
035 $a (MiAaPQ)AAI31140321
035 $a AAI31140321
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Davies, Michael E. $3 1476591
245 1 0 $a Composable Architecture Primitives for the Era of Efficient Generalization.
264 0 $c 2024
300 $a 1 online resource (143 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
500 $a Advisor: Sankaralingam, Karthikeyan.
502 $a Thesis (Ph.D.)--The University of Wisconsin - Madison, 2024.
504 $a Includes bibliographical references
520 $a We are in the age of AI, with rapidly changing algorithms and hardware. As Deep Learning (DL) applications have evolved, architects have turned to specialization to keep up with the insatiable compute that DL demands. Counter to this hardware evolution, DL demands that hardware support a vast number of diverse application behaviors and in addition, an inflexible ossified software interface in order to be viable. This dissertation motivates and adopts the guiding philosophy of "Efficient Generalization" - as opposed to "Specialization" - as a pathway for architects to support diverse DL application behavior while also catering to the ossified DL stack. We take a principled approach to identifying architecture and software primitives to enable performance and efficiency growth while paying careful attention to the restrictions that come with an ossified software stack and reaching the limits of technology scaling. We calibrate our understanding of modern architecture for DL by conducting a longitudinal study of how the DL revolution has panned out for GPUs - the dominant AI hardware platform of the day. Based on the insights from this study, we explore the fundamental architecture primitives needed to support DL's conflicting needs of programmability and high-performance. We use the knowledge of these architecture primitives to then tease apart the diminishing impacts of technology scaling. Combining all our findings about the role of technology, microarchitecture and architecture, we explore a novel execution model, spatial fusion, for DL on existing programmable architectures (GPUs) which can enable better utilization of on-chip resources leading to higher performance and power efficiency with only modest changes to the architecture.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Computer science. $3 573171
650 4 $a Engineering. $3 561152
653 $a Deep Learning
653 $a AI hardware
653 $a Power efficiency
653 $a Technology scaling
653 $a Software stack
655 7 $a Electronic books. $2 local $3 554714
690 $a 0984
690 $a 0800
690 $a 0537
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a The University of Wisconsin - Madison. $b Computer Sciences. $3 1179878
773 0 $t Dissertations Abstracts International $g 85-11B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31140321 $z click for full text (PQDT)