Processing-in-Memory for AI: From Circuits to Systems

168 Pages·2022·7.484 MB·English

by  Joo-Young Kim, Bongjin Kim, Tony Tae-Hyoung Kim

Joo-Young Kim
Bongjin Kim
Tony Tae-Hyoung Kim   Editors
Processing-in-
Memory for AI
From Circuits to Systems
Processing-in-Memory for AI
Joo-Young Kim • Bongjin Kim
Tony Tae-Hyoung Kim
Editors
Processing-in-Memory
for AI
From Circuits to Systems
Editors
Joo-YoungKim BongjinKim
KoreaAdvancedInstituteofScienceand UniversityofCalifornia
Technology SantaBarbara,CA,USA
Daejeon,Korea(Republicof)
TonyTae-HyoungKim
NanyangTechnologicalUniversity
Nanyang,Singapore
ISBN978-3-030-98780-0 ISBN978-3-030-98781-7 (eBook)
https://doi.org/10.1007/978-3-030-98781-7
©TheEditor(s)(ifapplicable)andTheAuthor(s),underexclusivelicensetoSpringerNatureSwitzerland
AG2023,correctedpublication2023
Thisworkissubjecttocopyright.AllrightsaresolelyandexclusivelylicensedbythePublisher,whether
thewholeorpartofthematerialisconcerned,specificallytherightsoftranslation,reprinting,reuse
ofillustrations,recitation,broadcasting,reproductiononmicrofilmsorinanyotherphysicalway,and
transmissionorinformationstorageandretrieval,electronicadaptation,computersoftware,orbysimilar
ordissimilarmethodologynowknownorhereafterdeveloped.
Theuseofgeneraldescriptivenames,registerednames,trademarks,servicemarks,etc.inthispublication
doesnotimply,evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfromtherelevant
protectivelawsandregulationsandthereforefreeforgeneraluse.
Thepublisher,theauthorsandtheeditorsaresafetoassumethattheadviceandinformationinthisbook
arebelievedtobetrueandaccurateatthedateofpublication.Neitherthepublishernortheauthorsor
theeditorsgiveawarranty,expressedorimplied,withrespecttothematerialcontainedhereinorforany
errorsoromissionsthatmayhavebeenmade.Thepublisherremainsneutralwithregardtojurisdictional
claimsinpublishedmapsandinstitutionalaffiliations.
ThisSpringerimprintispublishedbytheregisteredcompanySpringerNatureSwitzerlandAG
Theregisteredcompanyaddressis:Gewerbestrasse11,6330Cham,Switzerland
Contents
1 Introduction .................................................................. 1
Joo-YoungKim
2 Backgrounds.................................................................. 15
ChengshuoYu,HyunjoonKim,BongjinKim,
andTonyTae-HyoungKim
3 SRAM-BasedProcessing-in-Memory(PIM).............................. 41
HyunjoonKim,ChengshuoYu,andBongjinKim
4 DRAM-BasedProcessing-in-Memory..................................... 65
DonghyukKimandJoo-YoungKim
5 ReRAM-BasedProcessing-in-Memory(PIM)............................ 93
TonyTae-HyoungKim,LuLu,andYuzongChen
6 PIMforMLTraining........................................................ 121
JaehoonHeoandJoo-YoungKim
7 PIMSoftwareStack ......................................................... 143
DonghyukKimandJoo-YoungKim
8 Conclusion.................................................................... 161
Joo-YoungKim,BongjinKim,andTonyTae-HyoungKim
Correctionto:Processing-in-MemoryforAI ................................. C1
v
Chapter 1
Introduction
Joo-YoungKim
1.1 HardwareAccelerationforArtificialIntelligenceand
MachineLearning
Artificialintelligence(AI)andmachinelearning(ML)technologyenablecomputers
to mimic the cognitive tasks believed to be what only humans can do, such as
recognition, understanding, and reasoning [1]. A deep ML model named AlexNet
[2],whichuseseightlayersintotal,wonthefamouslarge-scaleimagerecognition
competition called ImageNet by a significant margin over shallow ML models in
2012. Since then, deep learning (DL) revolution has been ignited and spread to
many other domains such as speech recognition [3], natural language processing
[4],virtualassistance[5],autonomousvehicle[6],androbotics[7].Withsignificant
successes in various domains, DL revolutionizes a wide range of industry sectors
suchasinformationtechnology,mobilecommunication,automotive,andmanufac-
turing[8].However,asmoreindustriesadoptthenewtechnologyandmorepeople
useitdaily,wefaceanever-increasingdemandforanewtypeofhardwareforthe
workloads.ConventionalhardwareplatformssuchasCPUandGPUarenotsuitable
for the new workloads. CPUs cannot cope with the tremendous amount of data
transfers and computations required in the ML workloads, while GPUs consume
largeamountsofpowerwithhighoperatingcosts.
AI chip or accelerator is the hardware that enables faster and more energy-
efficient processing for AI workloads (Fig.1.1). Over the past few years, many
AI accelerators have been developed to serve the new workloads, targeting from
battery-powerededgedevices[9–11]todatacenterservers[12].AsMcKinseypre-
dictedinthereport[13],theAIsemiconductorindustryisexpectedtogrow18–19%
J.-Y.Kim((cid:2))
SchoolofElectricalEngineering(E3-2),KAIST,Daejeon,SouthKorea
e-mail:[email protected]
©TheAuthor(s),underexclusivelicensetoSpringerNatureSwitzerlandAG2023 1
J.-Y.Kimetal.(eds.),Processing-in-MemoryforAI,
https://doi.org/10.1007/978-3-030-98781-7_1
2 J.-Y.Kim
Fig.1.1 AIchipanditsmarketprediction
every year to 65 billion, accounting for about 19% in the entire semiconductor
market in 2025. So far, the AI hardware industry is led by big tech companies.
Googledeveloped theirownAIchipnamedtensorprocessingunit(TPU)thatcan
workwithTensorFlow[14]softwareframework.AmazondevelopedInferentiachip
[15] for high-performance ML inference. Microsoft’s BrainWave [16] uses FPGA
infrastructure to accelerate ML workloads at scale. Even an electric car maker
Tesla developed the full self-driving (FSD) chip for autonomous vehicles. There
are many start-up companies in this domain. Habana Labs, acquired by Intel in
late 2019, developed Gaudi processor for AI training. Graphcore has developed
intelligent processing unit (IPU) [17] and deployed in datacenters. Groq’s tensor
streamingprocessor[18]optimizesdatastreamingandcomputationswithfixedtask
scheduling.Cerabras’swafer-scaleengine[19]triestouseawholewaferasaML
processortokeepalargemodelwithoutexternalmemories.
1.2 MachineLearningComputations
In this section, we introduce the basic models of deep neural networks (DNNs)
andtheircomputations.ADNNmodeliscomposedofmultiplelayersofartificial
neurons, where neurons of each layer are inter-connected with the neurons in the
neighborlayers.ThemathematicalmodeloftheneuroncomesfromFrankRosen-
blatt’s Perceptron [20] model, as shown in Fig.1.2. Inspired by the human neuron
model, it receives multiple inputs among many input neurons and accumulates
their weighted sums with a bias. Then it decides the output through an activation
function, where the activation function is non-linear and differentiable, having a
step-likecharacteristicshape.Asaresult,theoutputofaneuronisexpressedwith
thefollowingequation:
1 Introduction 3
Fig.1.2 Artificialneuron:perceptronmodel
Fig.1.3 Fullyconnectedlayer
y =f(w x +w x +...+w x +b) (1.1)
1 1 2 2 n n
Based on the network connection, there are three major layers in the DNN
models,whichdeterminestheactualcomputations:fullyconnected,convolutional,
andrecurrentlayer.
1.2.1 FullyConnectedLayer
Figure1.3showsthefullyconnectedlayerthatinterconnectstheneuronsintheinput
layertotheneuronsinthenextlayer.Theinputvectoristhevaluesofinputneurons,
a3×1vectorinthiscase,andtheoutputvectoristhevaluesofoutputneurons,a
4×1vector.Eachconnectionofthefullyconnectednetworkbetweenthetwolayers
representsaweightparameterinthemodel.Forexample,W representsaweight
01
parameter of the connection between the input neuron 0 and the output neuron 1.
Collectively,thenetworkbecomesa4×3weightmatrix.Inaddition,eachoutput
neuronhasabias,sothelayerhas4×1biasvector.Foreachoutputneuron,wecan
4 J.-Y.Kim
Fig.1.4 Convolutionallayer
writetheoutputvaluey usingEq.1.1.Asaresult,theequationscanbeformulated
intoamatrix-vectorequationasfollows:
y =f(Wx+b) (1.2)
Ifthemodelincludesmultiplelayers,whichisthecaseofdeepneuralnetworks,the
matrix operations will be cascaded one by one. Traditional multi-layer perceptron
(MLP)modelsaswellasthelatesttransformermodels[21]arebasedonthefully
connectedlayer.
1.2.2 ConvolutionalLayer
Theconvolutionallayeriterativelyperforms3-dconvolutionoperationsontheinput
layer using multiple weight kernels to generate the output layer, as illustrated in
Fig.1.4.Theinputlayerhasmultiple2-dinputfeaturemaps,sizedH× H× C,and
thesizeofeachkernelisK× K× C.Forcomputation,itperforms3-dconvolution
operationsfromtop-lefttobottom-rightforeachkernelwithastrideofU.Asingle
convolutionoperationaccumulatesalltheinnerproductsbetweentheinputandthe
kernel.Asaresultofscanningforakernel,itgetsasingleoutputfeaturemapsized
E× E.Byrepeatingthisprocessforallkernels,theconvolutionallayerproduces
thefinaloutputlayer,sizedtoE× E× M.Theequationforanoutputpointinthe
convolutionallayerisasfollows:
1 Introduction 5
Fig.1.5 Recurrentlayer
C(cid:2)−1K(cid:2)−1K(cid:2)−1
O[u][x][y]=B[u]+ I[k][Ux+i][Uy+j]W[u][k][i][j],
k=1 i=1 j=1 (1.3)
(H −R+U)
0≤u<M,0≤x,y <E,E =
U
Many convolution neural network (CNN) models use a number of convolutional
layers with a few fully connected layers at the end for image classification and
recognitiontask[22].
1.2.3 RecurrentLayer
Figure 1.5 shows the recurrent layer that has a feedback loop from the output to
the input layer in the fully connected setting. In this layer, the cell state of the
previous timestamp affects the current state. Its computation is also matrix-vector
multiplicationbutinvolvesmultiplestepswithdependency.Itscellandoutputvalue
are expressed as follows. The hyperbolic tangent is usually used for activation
functionintherecurrentlayer.
ht =f(Uhxt +Vhht−1+bh),Ot =f(Whht +bo) (1.4)
Recurrentneuralnetworks(RNNs)suchasGRUs[23]andLSTMs[24]arebased
onthistypeoflayerandpopularlyusedforspeechrecognition.