暂无图片
暂无图片
暂无图片
暂无图片
暂无图片
图神经网络加速结构综述 - 李涵, 严明玉, 吕征阳, 李文明, 叶笑春, 范东睿, 唐志敏.pdf
95
26页
0次
2021-11-16
免费下载
DOI
:
issn
JournalofCom
p
uterResearchandDevelo
p
ment
(
):
,
 
稿
:
;
:
 
:
(
,
,
);
(
KYSB
);
(
A
)
This work wassu
pp
orted b
y
the National NaturalScience Foundation of China
(
,
,
),
the
InternationalPartnershi
p
Pro
g
ramofChineseAcadem
y
ofSciences
(
KYSB
),
andtheO
p
enPro
j
ectPro
g
ramofthe
StateKe
y
Laborator
y
ofMathematicalEn
g
ineerin
g
andAdvancedCom
p
utin
g
(
A
)
 
:
(
y
anmin
gy
uictaccn
)
 
,
 
,
 
,
 
 
 
,
 
,
(
(
)
 
 
)
(
 
 
)
(
lihanamsictaccn
)
Surve
y
onGra
p
hNeuralNetworkAccelerationArchitectures
LiHan
,
,
Yan Min
gy
u
,
,
LüZhen
gy
an
g
,
,
LiWenmin
g
,
YeXiaochun
,
FanDon
g
rui
,
,
and
Tan
g
Zhimin
,
(
StateKe
y
Laborator
y
o
f
Com
p
uterArchitecture
(
Instituteo
f
Com
p
utin
g
Technolo
gy
,
ChineseAcadem
y
o
f
Sciences
),
Bei
j
in
g
)
(
Universit
y
o
f
ChineseAcadem
y
o
f
Sciences
,
Bei
j
in
g
)
Abstract Recentl
y
,
theemer
g
in
gg
ra
p
hneuralnetworks
(
GNNs
)
havereceivedextensiveattention
fromacademiaandindustr
y
duetothe
p
owerful
g
ra
p
hlearnin
g
andreasonin
g
ca
p
abilities
,
andare
consideredtobethecoreforcethat
p
romotesthefieldofartificialintelli
g
enceintothe
co
g
nitive
intelli
g
ence
sta
g
eSinceGNNsinte
g
ratetheexecution
p
rocessofbothtraditional
g
ra
p
h
p
rocessin
g
andneuralnetwork
,
ah
y
bridexecution
p
atternnaturall
y
exists
,
which makesirre
g
ularandre
g
ular
com
p
utation and memor
y
access behaviors coexist.This execution
p
attern makes traditional
p
rocessorsandtheexistin
gg
ra
p
h
p
rocessin
g
andneuralnetworkaccelerationarchitecturesunableto
co
p
ewiththetwoo
pp
osin
g
executionbehaviorsatthesametime
,
andcannotmeettheacceleration
re
q
uirementsofGNNsTosolvetheabove
p
roblems
,
accelerationarchitecturestailoredforGNNs
continuetoemer
g
eThe
y
customizecom
p
utin
g
hardwareunitsandonchi
p
stora
g
elevelsforGNNs
,
o
p
timizecom
p
utationandmemor
y
accessbehaviors
,
andhaveachievedaccelerationeffectswellBased
onthechallen
g
esfaced b
y
the GNN accelerationarchitecturesinthedesi
g
n
p
rocess
,
this
p
a
p
er
s
y
stematicall
y
anal
y
zes and introduces the overall structure desi
g
n and the ke
y
o
p
timization
technolo
g
iesin thisfield from com
p
utation
,
onchi
p
memor
y
access
,
offchi
p
memor
y
access
res
p
ectivel
y
Finall
y
,
thefuturedirectionofGNN accelerationstructuredesi
g
nis
p
ros
p
ectedfrom
differentan
g
les
,
anditisex
p
ectedtobrin
g
certainins
p
irationtoresearchersinthisfield
Ke
y
words 
g
ra
p
h neuralnetwork
;
h
y
bridexecution
p
attern
;
accelerationarchitecture
;
artificial
intelli
g
ence
;
domains
p
ecificarchitecture
 
 
,
,
广
,
,
访
,
,
,
,
访
,
,
访
访
,
 
;
;
;
;
 TP
  
,
(
convolu
tionalneuralnetworks
,
CNNs
)
(
recurrentneuralnetworks
,
RNNs
)
(
Euclideans
p
ace
)
[
]
,
,
(
线
)
[
]
,
,
,
,
(
g
ra
p
hneuralnetworks
,
GNNs
)
使
,
广
[
]
[
]
[
]
[
]
,
)
,
,
,
访
)
,
(
multila
y
er
p
erce
p
trons
,
MLPs
)
,
访
,
访 使
使
CPU
GPU
线
[
]
,
H
y
GCN
[
]
,
使
,
,
访
访
,
[
]
,
[
]
,
,
线
[
]
(
re
p
resentationlearnin
g
)
,
 
:
of 26
免费下载
【版权声明】本文为墨天轮用户原创内容,转载时必须标注文档的来源(墨天轮),文档链接,文档作者等基本信息,否则作者和墨天轮有权追究责任。如果您发现墨天轮中有涉嫌抄袭或者侵权的内容,欢迎发送邮件至:contact@modb.pro进行举报,并提供相关证据,一经查实,墨天轮将立刻删除相关内容。

评论

关注
最新上传
暂无内容,敬请期待...
下载排行榜
Top250 周榜 月榜