In this paper, we propose a cross-layer algorithm/architec-ture co-exploration for wireless multimedia systems to coor-dinate interactions among sub-layer optimizers for improve-ments in energy/QoS/reliability. By exploiting the inherentredundancy in wireless multimedia systems, we generate anexpanded design space over traditional, layer-specific ap-proaches. Specifically, we control the error resilient encoderat the application layer to provide awareness of architecturalexploration at the physical layer (WCDMA) allowing newdesign points with lower power consumption via aggressivevoltage scaling. While trying to reduce energy consump-tion, the fault tolerant technique compensates the effect ofthe hardware and network errors due to aggressive voltagescaling and lossy transmission, respectively. Our experi-ments on H.263 video over WCDMA communication systemdemonstrate that co-exploration enlarges the feasible designspace which results in significant power savings of 34% inthe H.263 encoder more than 20% in the WCDMA modem.
1. INTRODUCTION
Mobile multimedia applications on battery-powered devicesrunning typically operate under harsh wireless conditions, re-quiring good strategies for energy/QoS provisioning. This re-quires resource management policies at different layers in-cluding: a specific video encoding/decoding algorithm at theapplication layer; data transfer protocol and network moni-toring at the middleware layer; Fault Tolerant technique (FT)at the physical layer. Whereas traditional approaches con-sidered policies locally at each abstrated layer, out researchexpands the design space for energy/QoS by investigating op-portunities for cross-layer co-exploration. As shown in Fig-ure 1, without loss of generality, we consider a case study of ahypothetical wireless video communication platform. At theapplication layer, we use PBPAIR [1] energy-efficient error-resilient encoding. At the middleware layer, selective protec-tion [2] and UDP-Lite bit-error resilient packetization [3] isutilized. Lastly, we explore aggressive dynamic voltage scal-ing (AVS) on WCDMA memory [4, 5] at the physical layer.
Policies made at one layer can affect behavior at otherlayers. For example, if AVS is used at the physical layer(WCDMA), errors gradually start to appear in the hardware.
Specifically,logic. Theseand manifestin Figure 2 thrors (due to vsignal to noEach line deprotection (aClass-2 protTCOEF andFrom the saClass-1 andthe H.263 eerror free traof datagramof datagram.
For this erate (BER) llayers’ policthan 10−4 referent impacout that the award compostill remainsincreasing Pupper layerFor instancesider the BEBER is 10−4
There hations for mu
CrosCo−
Proto
WCD
CodeNetw
Phys
Appl
Amin Khajeh, Minyoung Kim, Nikil Dutt, Ahmed MUniversity of California, Irvine, CA
{akhajehd, minyounk, dutt, aeltawil, ku
Cross-Layer Co-Exploration of Exploiting EOver Wireless Applicati
Authorized licensed use limited to: Univ of Calif Irvine. Downlo
embedded memory is typically affected beforeerrors will then propagate through the system
themselves at the application layer, as illustratedat shows the impact of the WCDMA physical er-oltage scaling) on the quality of the video (peak
ise ratio, PSNR) for a sample video sequence.monstrates the effect of PBPAIR with selectivepplication/middleware layer). Class-1 and andection ensures bit-error free transmission exceptInter-TCOEF of H.263 payload [6], respectively.mple video sequence, we observe that data forClass-2 protection takes up to 25% and 80% ofncoded stream, respectively. To guarantee bit-nsmission, UDP-Lite packetization protects partwhere WCDMA cannot perform AVS on that part
xample, Figure 2 shows that WCDMA bit erroress than 10−5 is acceptable regardless of othery. On the other hand, WCDMA BER greatersults in undesirable quality of service with dif-t due to selective protection. It should be pointedbove results are generated from the straight for-sition of existing individual techniques. Thereopportunities for a larger design space such as
SNR by manipulating algorithmic parameters ofto accommodate physical layer’s characteristic., we can manipulate PBPAIR parameter to con-R, which in turn leads to acceptable PSNR when.
ve been several studies on cross-layer optimiza-ltimedia [7, 8, 9, 10]. In [7, 8], the authors have
Cognitive Radio Platforms [5]Power Management for
s−LayerExploration
col Design / Monitor
MA Control
c Parameter Tuning Encoding PBPAIR [1]Energy−efficient Error−resilient
ork Layer
ical Layer
ication Layer
Fig. 1. Cross-Layer Co-Exploration
. Eltawil, Fadi J. Kurdahi 92697, USArdahi}@uci.edu
rror Resilience for Video ons
ESTIMedia 2008
aded on October 1, 2009 at 19:08 from IEEE Xplore. Restrictions apply.
0
5
10
15
20
25
30
35
40
45
PS
NR
(dB
)
WCDMA Bit Error Rate
Protection Impact
10E-1 10E-2 10E-3 10E-4 10E-5 10E-6 Error Free
No protectionClass-1 protectionClass-2 protection
Fig. 2. The Initial Design Space
identified interactions between the different layers followedby coordinated optimization. They specifically explore mid-dleware and OS scheduling policies to control the overall en-ergy consumption without architectural exploration, whereaswe specifically study cross-layer co-exploration. The authorsof [9, 10] study the issues of cross-layer optimization as a newparadigm for network architecture to make better use of net-work resources. Those efforts are, however, mainly focusedon the architectural decisions in networking, not tuning thesystem parameters for QoS-energy optimization.
The main contribution of our work is algorithm/architectureco-exploration using cross-layer optimizations to develop areconfigurable video coding and processing system. Thekey idea underlying the co-exploration is that we manipulatethe parameters of the reconfigurable video coding to pro-vide awareness of architectural characteristics at the physicallayer. This paper contains the following specific contribu-tions: Comprehensive analysis of cross-layer QoS-energytradeoffs and coordinated interaction enable us to tune videocoding parameters on highly resource limited devices. Inour chosen H.263 video communication application andWCDMA simulation environment, our approach enlargesthe feasible design space by exploiting error resiliency due tothe redundant nature of video coding.
2. SUB-LAYER POWER-AWARE TECHNIQUES
We now illustrate sub-layer power-aware techniques for thethree layers shown in Figure 1.
2.1. Physical LayerWe consider WCDMA at the physical layer. The authorsin [4, 5] have shown that in some systems such as wirelessand multimedia, the computational engine does not need tobe 100% correct, 100% of the time. In fact, an insightfulview of wireless systems shows that these systems, by design,can accommodate a large degree of channel induced errorswhile meeting the stated system requirements such as targetBit Error Rate (BER). Therefore, by merging the system de-sign with the circuit design, a whole new design space can be
Fig. 3. Block
ior under Ag
explored wha similar manoise floor wapproach, thconditions tothe now extethe design, ners.
It is a wthe most effHowever, inded memoricontroller untaneously inthe embeddefaster than lo
In the prnique, we sethe rest of ththat is inhereories in the mtion of the dhave multiplcoding redunmunication nences a relatthe modulatiwe use this rcontrollable
In the Asupplied fromsupplied in a
14
Authorized licensed use limited to: Univ of Calif Irvine. Downlo
H.263(Decoder)
WCDMA(Rx)
Video (out)
AVS
�
ErrorResilient
Memories….….
Diagram of AVS WCDMA and ERM’s Behav-
gressive Voltage Scaling
ere controlled hardware errors can be treated innner to channel errors, thus contributing to thehile still meeting stated system metrics. In this
e key idea is to intentionally vary the operatinga point where errors start to occur and to exploitnded design space to optimize other aspects ofamely power consumption across different lay-
ell known fact that supply voltage reduction isective means of reducing power consumption.traditional voltage scaling techniques embed-
es are the bottleneck and typically the powerit reduces both voltage and frequency simul-order to maintain the correct functionality of
d memory (in general embedded memories failgic under aggressive voltage scaling).
oposed Aggressive Voltage Scaling (AVS) tech-parate the error-resilient memories (ERM) frome system. ERMs are memories that store datantly redundant such as raw data buffering mem-odem. In general, ERMs consume a large por-
esign area and they store raw soft bit values thate levels of redundancy. At the algorithmic level,dancy exists to protect the data against the com-oise. Also most of the time, the receiver experi-
ively higher SNR than the minimum required foron. Therefore, to achieve more power savingsedundancy and slack to allow some limited anderrors to occur in hardware.
VS technique, The error-resilient memories area variable source while the rest of the system in
traditional manner (i.e., nominal supply). Dur-
aded on October 1, 2009 at 19:08 from IEEE Xplore. Restrictions apply.
ing operation, the voltage to the ERMs is aggressively low-ered and errors are allowed to occur in the system, that willbe corrected at the application layer. In this paper, we exploresuch an approach to a cross-layer system that merges both themodem and the source coding in one system. A variable sup-ply (could be on chip or off-chip) supplies the ERMs while astandard supply provides the rest of the Modem with nominalVdd. The authors are aware that this is one example, and thatthe techniques of error resilience as well as the power savingare design dependent. However, the main point of the paper isthat that it is possible to utilize the knowledge of the require-ments of the application using the modem to reduce power.
Figure 3 shows the block diagram of the receiver whichutilizes the AVS technique. The received signal goes intothe WCDMA receiver which is equipped with an AVS con-troller. Based on the operating condition (i.e., bit rate and thereceived signal’s SNR and etc.), the controller decides on theproper voltage for the ERMs. It is very important to quan-tify the effect of aggressive voltage scaling on the ERM’sfunctionality. To do so, we set up a HSpice simulation on6-Transistor Static random access memory (SRAM) in 65nmtechnology using Predictive Technology Model (PTM) [11]and statistically calculated the probability of failure (ReadAccess Failure, Write Failure and Destructive Read Failure)for voltages lower than nominal (VddNominal
= 0.9v) [4]. Thegraph under the Error-Resilient-Memories in Figure 3 showshow the ERM’s behave under AVS. As the result of processvariation, the errors are gradually increasing in log-domainwith the voltage reduction.
2.2. Network Layer
The network layer attempts to monitor the current networkstatus and control the data transmission based on the networkstatus. The impact of transmission errors on video quality de-pends on the spatial and temporal location of the error. Forinstance, errors in packet headers or motion vectors can causedrastic video quality degradation. A single bit error can dam-age a major part of a frame due to resynchronization of vari-able length codes. For the errors propagated among consec-utive P-frames, intra coding (that we will discuss in the fol-lowing subsection) can stop error propagation in temporal do-main.
In this context, protecting the entire bitstream is ex-pensive because of bandwidth limitations and delay con-straints. Selective protection on the most critical information[2] combined with bit-error resilient packetization schemefor UDP-Lite [3] can be effectively used for this purpose.In this particular work, we assume selective protection andUDP-Lite implementation are available and only focus onthe co-exploration of the other two layers (application andphysical layer).
2.3. Application Layer
We consider PBPAIR (Probability Based Power Aware IntraRefresh) [1] as an application layer technique. The PBPAIR
frame
Sta
Pro
ceed
wit
h n
ext
fram
escheme inserany other frbitstream atimproves errencoding eneestimation (in a predictiPBPAIR utiltempts to allmacro blockthereby prodseen in Figu
The remate MBs to mconsumptiontwo parametold: IntraThand raw videbustly encodillustrates simre-evaluateslustrated asencoding moand to find m5). The encotween probavalue (IntraTencoded as ialready expethat point anerror resiliencoded as a inon a heuristiage content i
15
Authorized licensed use limited to: Univ of Calif Irvine. Downlo
time
frameloss
1 frame2 frame3 frame4 frame5
Fig. 4. Partial Intra-coding
Encode as Intra MB
Encoding modeselection
rt from error free image frame
P
IME
SAD check
Yes
No
Encode as Inter MBLast MB?
Update info.& go to next frame
Yes
No
Fig. 5. PBPAIR Overview
ts intra-coding (i.e., coding without reference toame) to enhance the robustness of the encodedthe cost of compression efficiency. Intra-codingor resilience, but it also contributes to reducingrgy consumption since it does not require motion
which is the most power consuming operationve video compression algorithm). In particular,izes partial intra-coding. Partial intra-coding at-eviate the error propagation by using intra-coded(MB) — the burden of refreshing is distributed,ucing a much smoother output rate — as can bere 4.
aining issue is how we pick the most appropri-aximize error resiliency while reducing energy
. For this purpose, the PBPAIR scheme takesers i) the user’s QoS expectation (Intra Thresh-), and ii) the network packet loss rate (PLR: α),o sequences as inputs to generate a bitstream ro-ed against network transmission errors. Figure 5
plified version of PBPAIR algorithm. PBPAIRthe probability of correctness of each MB (as il-Encoding Mode Selection in Figure 5) to decidede (I:Intra-coded MB, P:Predictively-coded MB)otion vector (ME:Motion Estimation in Figure
ding mode selection is done by comparison be-bility of correctness of a MB (σ) and a thresholdh). A MB with lower σ than IntraTh should bentra MB (refresh) since that particular MB hasrienced a sufficient amount of inter-coding up tod IntraTh values can be considered as requestedcy level. For a MB that is determined to be en-ter-coded macro block, motion estimation basedc that considers both network condition and im-s performed.
aded on October 1, 2009 at 19:08 from IEEE Xplore. Restrictions apply.
It should be pointed out that the parameters can be easilymanipulated to cope with other layers’ operating condition.For example, PBPAIR increases intra-coding by lowering theIntraTh parameter when there is high network packet loss(monitored at network layer). Indeed, PBPAIR controls thecoding efficiency, error resiliency, and power consumption bytuning these parameters to achieve cross-layer co-exploration.We explain these details in our approach (Section 3).
3. CO-EXPLORATION
In this section, we discuss how to coordinate the individualtechniques (physical layer and application layer) in a cross-layer manner based on the operating condition. Figure 6shows the overall system model as an exemplar for our cross-layer co-exploration approach. As mentioned in the previoussection, we assume a perfect packetizer e.g., UDP-Lite [3]for selective protection. To do so we need to:
• quantify the effect of memory errors at different bit rateon WCDMA bit error rate (BER), and
• explore the methods of taking the WCDMA BER intoaccount and be compensating for at the applicationlayer.
In the WCDMA system, the data buffering and de-interleaving memories (Error Resilient Memories) consumeapproximately 50% of the overall memory required for theentire modem. Furthermore, two instances of the ERMs aloneaccount for 45% of the total on-chip memory power. Thesetwo instances are used to buffer the received data right afterRAKE combining and to store the de-interleaved symbolsprior to rate matching and processing by the decoder. Inour simulation, we used the memory model error that wedescribed in Section 2.1 for these two memories and sim-ulated the WCDMA end-to-end physical layer for differentbit rates. Figure 7 shows the effect of the memory errors onthe WCDMA BER for different transmission bit rates. Asexpected, given the same SNR, for higher bit rates, the mem-ory errors have higher impact on the BER since there is lessredundancy in the system.
Once we quantified the effect of ERM errors on WCDMAbit error rate, We need to pass this information to the applica-tion layer in order to adapt to it. First, we attempt simple com-position. As shown in Figure 6, the sender encodes raw video(application layer) then packetizes (network layer) and trans-mits encoded bitstream via WCDMA (physical layer) to thenetwork. At the application layer, PBPAIR generates a robustbitstream against monitored network packet loss rate, α. Atthe physical layer, WCDMA induces bit errors in the memoryat the receiver side to reduce power consumption via voltagescaling. We define this impact as β. (i.e., β = 1− (1− ber)n,where ber and n represent bit error rate and number of bits,respectively.) Under the condition of α network packet lossrate and β WCDMA induced error, the probability of errorin the path (encoding → transmission) can be calculated as
following:
As showure 6, in thisα′ to the oriagainst netwror impact (βsince the imptinuously incβ = 1 − (1to refresh thelayer awaren
Alternatitain β at thevalue of encblock with irectness) thanentially decwith α erroring mode seinequality:
A simple cal
where Nα inSince the
the refresh pcorrespondin
EncodedS
EncodedS
On the o(Qvalue) anthe regressio
where c1 ancharacteristi
Thereforthe new quanpropriate lev
Nα
Nα
∴
1The exact
between consec
in this particula2Finding the
c1 and c2 for th
previous work [
16
Authorized licensed use limited to: Univ of Calif Irvine. Downlo
α′ = 1− (1− α)× (1− β)
n in ths Application Layer of the Sender in Fig-simple composition, we feed this adjusted error
ginal PBPAIR instead of α to generate bitstreamork condition (α) as well as physical layer er-). Interestingly, this leads to an unstable systemact of bit error in WCDMA memory, β, will con-rease due to the larger encoded bitstream (i.e.,− ber)n with larger n) by inserting intra-coding
errors. This effect highlights the need for crossess to avoid such an undesired outcome.vely, a successful approach is to attempt to main-similar level by manipulating the quantization
oding. Recall that PBPAIR encodes a macrontra-coding if it has lower σ (probability of cor-n IntraTh (a given threshold). The σ value expo-reases by the probability of correct transmission(i.e., 1−α). In other words, the PBPAIR encod-lection is approximately1 based on the following
σ ≈ (1− α)Nα < IntraTh
culation leads to
Nα = � ln(IntraTh)ln(1− α)
�
dicates the refresh period.length of bitstream is inversely proportional to
eriod, we can calculate the ratio of encoded sizeg to the error.
izeα′
izeα≈ Nα
Nα′≈ ln(1− α′)
ln(1− α)= 1 +
ln(1− β)ln(1− α)
ther hand, we observe that a quantization valued encoded size has power relation according ton on profiled video stream like below:
EncodedSizeα = c1 ×Qvaluec2α
d c2 are constants depending on the video inputc2.e, given α, β, and Qvalueα, we can figure outtization value Qvalueα′ to maintain β at an ap-
el.
′≈ EncodedSizeα′
EncodedSizeα= (
Qvalueα′
Qvalueα)c2
Qvalueα′ = �Qvalueα × (Nα
Nα′)
1c2 �
equation in [1] also takes into account the similarity factorutive frames. However, we do not need that level of accuracy
r work.
constants c1 and c2 is out of scope of this paper. We estimate
e test video sequences, and the result is consistent with our
7].
aded on October 1, 2009 at 19:08 from IEEE Xplore. Restrictions apply.
(α,β)fα’
Phy.
App.
β
Aggressive DVS
Sender
Partial Intra Coding
Selective Protection
PBPAIR
UDP−Lite
WCDMA AVS Tx
Net. (Bit ErrorImpact)
LoNet
α (Packet Loss Rate
Fig. 6. System Model for Cross-Layer Co
10−5 10−4 10−3 10−2 10−110−8
10−6
10−4
10−2
100
Memory Error
WC
DM
A B
ER
144Kbps, SNR=5dB64Kbps, SNR=5dB64Kbps, SNR=4dB
Fig. 7. Effect of the WCDMA Memory Errors on the System
Bit Error Rate
When Qvalueα′ is out of bound (i.e., larger than 31 in caseof PBPAIR implementation), we consider that as an infeasibledesign point.
In summary, we considered cross-(Application and Physical)-layer co-exploration by exposing physical errors in theWCDMA up to the PBPAIR application layer. Specifically,we provide PBPAIR with α′ and Qvalueα′ to consider thenetwork packet loss (α) and the impact of WCDMA error (β)induced by AVS in a cooperative manner. In this section, weillustrated that individual optimization techniques (such asPBPAIR and AVS) or simple composition without consider-ing other layers’ characteristic may results in drastic qualitydegradation. In the following section, we will discuss thebenefit of power saving using this co-exploration.
4. EXPERIMENTS
We use a PBPAIR implementation from our previous work[1], and feed the new α′ instead of the PLR (α) of the originalPBPAIR. We assume that a simple copy scheme is used forerror concealment at the decoding side. Note that we use auniform distribution of frame discard to generate the packetloss pattern. For simplicity, but without loss of generality, weuse the frame loss rate to denote the network packet loss rate.For the encoding/decoding power consumption, we utilize theSimics [12] full system simulation platform, capable of simu-lating target systems that include real network connection andrun operating systems and workloads. Specifically, we use the
100%
5%
10%
15%
20%
25%
30%
35%
WC
DM
A P
ower
Sav
ings
Fig. 8. WCD
and Aggress
Simics modeLinux 2.4 to
Figure 8form introduused 65nm tedata bufferinure 3 one caory errors.age, dynamiclated. Class-80% of the mage) to guarashown in Figings.
Next, webenefit of cvideo qualityIt should befeasible desiis less than 15 design poiHowever, bycan expanderrors up toin Figure 9.resilient naturameter. Notenergy savin
17
Authorized licensed use limited to: Univ of Calif Irvine. Downlo
Data Receiving
Receiver
WCDMA Rx
De−packetizingUDP−Lite
H.263 DecodingTMNDEC App.
Net.
Phy.
ssywork
)
-exploration
−20 10−15 10−10 10−5 100
Memory Error
No ProtectionClass−1 ProtectionClass−2 Protection
MA Power Savings due to Selective Protection
ive Voltage Scaling (AVS)
l of a PowerPC-based Ebony card [13] that bootsestimate energy consumption.
shows the energy savings on the simulation plat-ced in [5]. For the power savings calculations wechnology. The x-axis represents the errors in theg memories of the WCDMA receiver. From Fig-n relate the memory supply voltage to the mem-Based on the value of the memory supply volt-
and leakage power of the memory can be calcu-1 and Class-2 protection corresponds to 25% andemory supplied with high voltage (nominal volt-ntee the bit-error free operation, respectively. Asure 8, more protection leads to less power sav-
present experimental results that illustrate theross-layer co-exploration. Figure 9 shows the
and energy consumptions of our co-exploration.pointed out that as mentioned in Section 1, thegn space for video streaming is the case when β0−4 (Figure 2), which corresponds to the bottomnts in Figure 7 for the 64Kbps and 5dB SNR.applying the proposed cross layer approach we
the design space to accommodate physical layer10−2 with minimal video degradation as shown
This expanded design space utilizes the errorre of enhanced PBPAIR encoding with α′ pa-
e that in this situation, we achieve more that 20%gs from WCDMA (Figure 8) without any quality
aded on October 1, 2009 at 19:08 from IEEE Xplore. Restrictions apply.
26
28
30
32
34
PS
NR
(dB
)
WCDMA Memory Error Rate
Video Quality
10E-1 10E-2 10E-3 10E-4 10E-5
No protectionClass-1 protectionClass-2 protection
(a)
25
25.5
26
26.5
27
27.5
Enc
odin
g E
nerg
y (J
)
WCDMA Memory Error Rate
Encoding Energy Consumption
10E-1 10E-2 10E-3 10E-4 10
(b)
Fig. 9. (a) Delivered Video Quality (in PSNR), (b) Encoding Energy Consum
Source: FOREMAN.QCIF 300frames, PLR: 10%)
loss. To further expand the design space to accommodatehigher physical layer error rates (10−1 WCDMA memory er-ror) PBPAIR with selective protection can be exploited whereprotection of Class-2 is assumed. As shown in Figure 9(a),the PSNR is improved at a cost of higher WCDMA memoryenergy (22% less energy reduction comparing with No pro-tection as shown in Figure 8). Decoding energy consumptionwill also increase by 3% as compared with No protectioncase3.
Based on these results at extremely high physical errorrates (10−1 WCDMA memory error) the power managementprotocol can select the following design points: (i) Class-2protection, (ii) No protection which results in 4% and 9% to-tal energy reduction at the cost of 1dB and 4dB PSNR degra-dation, respectively. Note that any power savings achieved atthe physical layer due to the proposed cross layer approachis in addition to the 17% to 34% encoding energy reduc-tion achieved at the application layer (as compared with othererror-resilient techniques) as documented in [1].
The novel aspect of our approach is that instead of opti-mizing each layer by itself, we try to reduce the power con-sumption of the overall system by passing on necessary in-formation between different layers and finding the optimumoperating parameters for the total system while meeting therequired metrics.
5. CONCLUSION AND FUTURE WORK
In this paper, we discussed the impact of algorithm/architectureco-exploration in the context of transmitting H.263 video overa wireless WCDMA modem. It is shown that cross-layerco-exploration by targeting data buffering memories in theWCDMA system and adaptively changing the data rate basedon the network information at the application layer, resultsin a considerable reduction in power consumption of 34% inthe H.263 encoder more than 20% in the WCDMA modem.The proposed approach expands the feasible design spaceallowing designers to explore larger tradeoffs during systemdesign.
3Encoding energy consumption does not depend on selective protection.
[1] Minyoun
Nalini V
error-res
intra ref
Rev., vol
[2] A. Nava
ing for m
[3] Jari Kor
tization
[4] Amin K
Rouwaid
voltage s
[5] Amin K
Fadi J. K
forms,”
[6] ITU-T R
trate Com
[7] S. Moha
R. Gupta
cross-lay
distribut
[8] Wangho
Jones, an
for mult
Mobile C[9] M. Chia
mathem
IEEE, vo
[10] S. Khan
optimiza
cost,” in
[11] Predictiv
edu/˜p[12] Simics F
simics[13] Ebony B
18
Authorized licensed use limited to: Univ of Calif Irvine. Downlo
E-5 3
3.1
3.2
3.3
3.4
3.5
Dec
odin
g E
nerg
y (J
)
WCDMA Memory Error Rate
Decoding Energy Consumption
10E-1 10E-2 10E-3 10E-4 10E-5
No protectionClass-1 protectionClass-2 protection
(c)
ption, (c) Decoding Energy Consumption (Video
6. REFERENCES
g Kim, Hyunok Oh, Nikil Dutt, Alex Nicolau, and
enkatasubramanian, “PBPAIR: an energy-efficient
ilient encoding using probability based power aware
