`
`A Portable Multimedia Terminal
`
`Successful personal communications terminals will depend upon
`the smooth integration of computation and communications
`facilities in a lightweight unit.
`
`Samuel Sheng, Anantha Chandrakasan, and R. W. Brodersen
`
` he personal communications indus-
`
`try has seen explosive growth in the past
`several years, especially in the number
`and types of services and technologies.
`In voiceband communications, systems
`such as mobile analog cellular tele-
`phony, radio pagers, and cordless telephones
`have become commonplace, despite their limited
`nature and sometimes poor quality of transmis-
`sion. In computing, portable “notebook” comput-
`ers are boasting capabilities far in excess ofthe desktop
`machines of five years ago, and multi-MIPS,
`RISC-based, portable workstations are available.
`Despite the myriad technologies to be had, how-
`ever, little integration of these diverse services —
`the combination of computation and communica—
`tions facilities in a portable unit — has occurred.
`Thus, our vision of a future personal communica-
`tions system (PCS) centers on such integration of
`services to provide ubiquitous access to data and
`communications via a specialized, wireless multi-
`media terminal.
`
`ROBERT W BRODERSEN
`is a member ofthe Electrical
`Engineering and Computer
`Science faculty of[he Univer-
`sity of California, Berkeley.
`
`transitory nature, such as stock pricing, local news,
`and so on, making distribution by other means such
`as CD-ROM impractical. Furthermore, given suf—
`ficiently large database servers, libraries of books,
`journal archives, and other currently “paper-inten-
`sive” media can be placed on—line; these databases
`would allow for instantaneous recovery of all types
`of information, without the need to be at a termi—
`nal physically attached to the wired network.
`Second, a PCS would have access to digital video
`databases containing both entertainment and
`educational media, such as animated information
`sequences, taped lectures, movies, news clips, and
`other isochronous data. Unlike today’s television
`broadcasts, video databases can be made avail-
`able on-demand, giving users the freedom to
`access video information as needed.Video data
`will be necessarily stored in a compressed format for
`minimization of both storage space and transmis-
`sion bandwidth, thus requiring that the wireless
`terminals at least support video decompression.
`Simplified entry mechanisms such as voice-recog-
`Aschcmatic view of such a system is shown in Fig.
`nition and handwriting-recognition interfacing to
`access the above functions would also be avail-
`l. The wireless terminal is in full duplex commu-
`nication with a networked base station, which serves
`able.The design ofan effective user interface to access
`as a gateway between the wired and wireless
`such a vast information storehouse is a critical
`mediums.Via a base station, users access services
`issue. By using speech recognition and pen—based
`over the high-speed communications backbone,
`input, supported by large, speaker-independent rec-
`including communicating with another person
`ognizers placed on the network, such interfacing and
`also linked into the network. This idea can also
`information access can be tremendously simpli-
`be extended to a user communicating not only
`fied. Placing the recognition units on the network
`with another person, but with network “servers.”
`conserves power in portable units and enables much
`Because the data bandwidth of future fiber-optic
`larger and more complex recognition algorithms
`networks is easily in excess of 10 Gb/s, these cen-
`to be employed. Recognizer servers can also make
`tralized servers can provide a wide variety of
`use of context-sensitive analysis, which can increase
`information services to users. A personal commu-
`recognition accuracy by determining which words
`nications system will likely include the four key
`are most likely to be used in a given application [1].
`features that follow.
`Fourth, the system would provide support for
`ANANTHA P. CHAN-
`Access to large commercial databases that con-
`a distributed computing environment, such as MIT’s
`DRAKASAN and SAMUEL
`tain information such as international and domes-
`SHENG are PhD. candi-
`X-Window system. In distributed computing envi-
`tic news, financial information, traffic data, trans-
`ronments, computation need not take place on a local
`dates in the Electrical Engi-
`portation schedules, voice mail, telephone numbers,
`machine; instead, computation is performed by pro—
`neering and Computer
`news, bulletin boards, and educational material is
`grams executing on one or more remote machines,
`Science department at the
`necessary. The continuous connectivity afforded
`which may have no computing capability except that
`University of California,
`by personal communications systems has several
`required to act as an intelligent display device. Many
`Berkeley.
`advantages. Many sources of information are of a
`such inexpensive “X~terminals” already exist. Unlike H.—
`64
`
`0163-6804/‘92/$03.00 1992© lEEE
`
`IEEE Communications Magazine - December 1992
`
`Petitioners - Exhibit 1010 Pagel
`
`Petitioners - Exhibit 1010 Page1
`
`
`
`
`
`
`High-speed fiber backbone
`
`
`
`
`Video
`
`database
`
`
`V Speech
`recognizer
`
`
`8-8
`
` Compute
`
`Compressed
`
`Commercial
`Video
`
`database
`(news, financial
`information, etc)
`
`biwee-
`
`Base
`station
`
`_
`
`
`
`
`
`Wireless multimedia
`Terminals
`- Xoterminal
`. Video display
`-' Audio input/output
`
`,
`
`
`
`
`
`TTY terminals that can only communicate with a
`single host machine, X—terminals possess all of the
`necessary networking capability to communicate
`with as many remote servers as needed. The multi—
`media terminal will be based on this model: remote
`computation servers will be used to run applications
`like spreadsheets, word processors, etc., with the re—
`sults being transmitted to the terminal. Likewise,
`supercomputer—class servers will perform intensive
`tasks requiring simulations, 3-D image rendering,
`and computer-aided design.
`Clearly, the cornerstone of the entire system lies
`in the ability of multimedia terminals and wire-
`less communications links to support all of the afore—
`mentioned services. Correspondingly, it is this desire
`for portability that translates directly into design con-
`straints on the size and weight of the terminals,
`the power they consume, and the frequency band-
`width needed in the wireless links.
`A diagram of a portable terminal is shown in
`Fig. 2. To minimize power, only those functions
`that are absolutely necessary are implemented:
`the analog RF transceiver; baseband processing
`for communications, such as equalization, coding, and
`paeketization; the image decompression unit and
`a display driver; and the speech codee. Because size
`is also an issue, a pen input system is integrated direct—
`ly onto a compact flat-panel display, eliminating the
`need for a large keyboard and providing greater visu—
`al feedback than possible with mouse- or track-
`ball-based interfaces. More important, the system is
`asymmetric in nature: High-quality, full—motion video
`is only supported in the downlink from the base
`station to the portable. This must be accounted
`for the in the design, as the bandwidth requirements
`in the reverse link from the portable unit are thus
`considerably less than the link from the base station.
`For video teleconferencing, a low—rate, reduced-
`quality video uplink might also be supported;
`however. the asymmetric bandwidth require-
`ments will still remain.
`The diagram shows that no direct user compu—
`tation is supported within the portable itself; instead,
`it is wholly dependent on the network servers to pro-
`vide desired functionality. Although this has imme-
`diate benefits in terms of reducing power con-
`sumption, it provides another advantage: Data
`that is highly sensitive to corruption will not be trans—
`mitted over the wireless network.
`Existing distributed computation environments
`are dependent on the fact that data transmitted over
`the network has high integrity — bit—error rates on
`wired Ethernet are typically on the order of l in
`1012 bits, and further protection is gained by pack-
`et retransmission after errors. On wireless networks,
`however, this is not true. Even after extensive appli—
`cation of error—correction coding, it is still diffi—
`cult to attain error rates even remotely as low as this.
`User “computation” data, such as spreadsheets or
`simulation results, simply cannot be allowed to
`sustain any corruption. For wireless systems, this
`translates into an inordinate amount oftransmission
`overhead in terms of coding and data retransmis—
`sion. On the other hand, user “multimedia” infor—
`mation, such as voice and image data, is relatively
`tolerant ofbit errors— an error in asingle video frame
`or an audio sample will not significantly change the
`Speech codee
`meaning or usefulness ofthe data. The ability to coex—
`
`ist with an error—prone transmission environment
`I Figure 2. Diagram of the portable wireless multimedia terminal.
`has tremendous impact on the overall system design.
`
`
`65
`
`
`
`servers
`
`A
`
`
`
`
`
`4 <7
`/
`
`*
`
`I Figure 1 . Personal communications system overview.
`
`
`
`Integrated pen input pad/
`flat panel display
`
`Antenna
`
`Microphone
`
`Battery
`
`Speaker
`
`
`
`
`
`
`RF transciever
`
`Packetizer/
`error correction
`
`Video decompression/
`display driver
`
`IEEE Communications Magazme - December 1092
`
`Petitioners - Exhibit 1010 Page2
`
`Petitioners - Exhibit 1010 Page2
`
`
`
`
`
`
` fl
`
`(a)
`
`overall time-average data rate. Minimizing over—
`all system bandwidth consumption while supporting
`a large number of users accessing data simultane-
`ously is paramount.
`One method to achieve this goal is to physical-
`ly reduce peak user data rates via data compression
`techniques (we will discuss this later). Another tech-
`nique, applied at the system level, is to utilize cel-
`lular networking techniques to achieve spatial fre-
`quency reuse. Because such a personal communi-
`cations system will first be uscd as a step beyond con-
`ventional wireless LANs, an indoor picocellular
`transmission environment will be of primary con-
`cern. (The techniques described in this article are
`applicable to both indoor and outdoor environments.)
`The advantages in improved spectral efficien-
`cy afforded by cellular systems have been employed
`extensively in present-day analog mobile radiotele-
`phony, where large-scale cells exploit these
`advantages to a limited extent. By scaling down
`cell sizes, tremendous increases in spectral efficiency
`canbe achieved. Asimple cellular scheme, as shown
`in Fig. 3a, consists of dividing the entire service
`area for the personal communication system into
`“cells” of radius R, with a single base station serving
`all mobile users within that cell. Each cell uses its
`own distinct set of frequencies. As users move
`from cell to cell, their transactions with the net-
`work are “handed off” from base station to base
`station, reconfiguring the network dynamically as
`the need arises.
`
`
`
`i
`
`2 "1 4
`\
`/
`/
`\
`\1,\3/—w1;
`\g/
`\
`/
`\ __/
`. 4 /__< Z
`>—
`\
`1 \
`,
`\
`/
`\_/
`
`‘
`.
`
`l
`
`(b)
`
`K=4
`D=3.46 R
`
`7
`
`,
`
`fl 6 H 5 >
`\
`/
`/
`1
`-\
`7 ,
`\
`1 )fi
`/_< 2 F—\
`\ 3
`\
`\\
`/
`\
`/
`\
`(MK 4 /H\ 3 // %\ 2 /
`>——/
`'> a 4 fl
`6
`5
`\
`/
`i
`/
`L < 7 ,H 6 )H’ 5 \>
`\_/
`\‘_< 7 /
`_/
`K=7 W
`D=4.58 R
`
`The key benefit of cellular systems is that they
`allow the network to achieve spatial multiple access.
`If two cells are separated by sufficient distance, each
`can use the same frequency bands at the same
`time without resulting in disastrous cochannel inter-
`ference. Thus, extensive frequency reuse becomes
`possible, as opposed to an umbrella scheme where
`every user must be assigned a different frequency
`slot. Fig. 3b shows several classical reuse patterns
`[2]; such patterns are typically characterized by a fre-
`quency reuse factor K, which represents the number
`of distinct frequency sets that need to be used to cover
`the entire service area. Instead of one user per
`frequency band, the network can now supportNusers
`per band, where N is the number of cells in the ser—
`vice area using that band. From the point of view of
`spectrum usage, each user effectively consumes only
`B/N Hz of bandwidth, where B is the physical band-
`width needed to support transmission, thus dras-
`tically increasing overall spectral efficiency.
`Clearly, minimizing the physical distance D be-
`tween cclls using the same frequency, by reducing
`the cell size R, yields the greatest frequency reuse,
`and hence the greatest gains in efficiency.1 There»
`fore, it is clear that the number of users support-
`able within the same overall system bandwidth in-
`creases quadratically as R decreases, because of the
`increased number ofcellular subdivisions within the
`service area. Minimizing R (and hence D) is criti—
`B ecause full-motion digital video is to be sup-
`ported, spectrum usage is of great concern ~
`cal in achieving high levels of spectral efficiency.
`per~user data rates can easily exceed 2 Mb/s even
`With an indoor environment, it is no longer fea-
`with the best compression schemes reported to date.
`sible to have only a single network transceiver sta-
`This data rate is not needed on a continuous basis;
`tion serving all of the terminals in the building. Due
`regular computation tasks such as word process-
`to the S to 15 dB attenuation through walls, the total
`ing or use of a spreadsheet require only slight screen
`microwave output power from all of the transmit-
`changes on a frame-by-frame basis over a small region,
`ters would have to be inordinately (and dangerously)
`usually on the order of a single character or a few
`high [3]. However, this attenuation can be taken
`related tofiandey
`pixels. Hence, it is probable that the peak data
`advantageofbyacellularnetwork—eachroom nat—
`D = R V3K
`rate required by users will be much larger than the
`urally becomes its own cell. Likewise, the cellular
`
`
`I Figure 3. (a) Cellular communications system; (b) typicalfi‘equency reuse
`patterns.
`
`Thus, the portable unit described above is truly a ter—
`minal dedicated to multimedia personal commu-
`nications, and not simply a notebook computer
`with a wireless LAN/modem attached to it.
`The remainder of this article will focus on sever~
`al of the major design issues behind portable mul»
`timedia terminals: spectrally efficient picocellular
`networking, low—power digital design, video data
`compression. and integrated wireless RF transceivers.
`Optimizing performance in each of these areas is
`crucial in meeting the performance requirements of
`the overall system and providing a small, lightweight
`terminal for personal communications.
`
`Picocell Networking and
`Spectrum Usage
`
`1 The frequency reuse dis-
`tance is geometrically
`
`66
`
`IEEE Communications Magazine ' December 1992
`
`Petitioners - Exhibit 1010 Page3
`
`Petitioners - Exhibit 1010 Page3
`
`
`
`
`
`
`
`RF
`Baseband
`
`tra nsceiver
`processing
`
`
`
`Equalization] L
`_
`-
`codi ng/packetizatioa:
`
`
`
`
`Multiplexer
`
`
`
`
`
`Audio D/A/
`speaker
`
`Microphone/
`speech codec
`
`”
`
`
`
`Video
`
`decompression
`
`‘
`
`_
`
`
`
`
`
`
`scheme now moves into three dimensions because
`the floors also provide RF isolation. Even if walls
`are not present, the use of electrically—adjustable
`directional antennas (such as a small phased-array
`device) can provide the same effect. The cells are now
`extremely small, on the order of about five meters;
`R is usually dictated by the size of the room, and Kcan
`be as low as 3 to 4, depending on how much atten—
`uation is provided by the walls. If K is increased to
`6 or 7, the assumption that cochannel interference
`is negligible becomes reasonable for most indoor
`office environments.
`In light of the above considerations, the total
`amount of spectrum that will be consumed to pro-
`vide the outlined services can now be addressed.
`After examination of the user density in a typical
`office environment,2 such as those found in mod—
`ern buildings with open-area soft-partition cubicles,
`cells with a radius of five meters typically contain
`12 to 16 active users. In the worst—case, a 2 Mb/s
`data rate for full-motion video using a linear DQPSK
`(differential quadrature phase-shift keying) mod-
`ulation scheme and design parameters from exist-
`ing systems [4-5], would require a transmission
`bandwidth of approximately 1.5 MHz per user using
`a 20 percent excess bandwidth raised-cosine
`pulse shape and a 25 percent loss for packetiza-
`tion, equalizer training, and other overhead. Assum-
`ing that half of the 16 users in the cell demand
`the complete 2 Mb/s data rate for full-motion
`digital video, and the remainder need 256 kb/s each3
`for lower data rate applications, a picocellular sys-
`tem with K = 7 would use approximately 100 MHz
`of bandwidth. Although 100 MHz is a consider-
`able amount of spectrum, this is amortized over large
`numbers of people using this spectrum simulta-
`neously within multiple buildings. Because the band-
`width of 100 MHz is designed to support full
`motion video and other multimedia network ser—
`vices for all users, this allocation of spectrum is
`not unreasonable, given the level ofservice provided
`bythe system, especiallywhen compared to the spec-
`trum allocated for existing systems such as NTSC
`television.
`There is another significant advantage to pico—
`cellular wireless systems: because transmit power
`is scaled down as cells move closer together to reduce
`interference, the power consumed in the portable’s
`transmitter to drive the antenna is corresponding-
`ly reduced. Whereas existing cellular systems use
`1 watt of transmit power for voiceband RF links
`in 5 mile cells, a picocellular system with 5 meter cells
`requires only milliwatts to maintain the link [6].
`
`
`
`
`
`I Figure 4. Signalflow diagram for the proposed multimedia terminal.
`
`Flat-panel display
`
`within the analog RF block, the receiver design
`becomes critical, because it must demodulatc a high—
`rate signal corrupted by noise and distortion, where-
`as the transmitter is relatively simple, with low
`data rate and output power requirements. Likewise,
`the algorithm chosen for the video decompres-
`sion should be designed, ifpossible, to make decom-
`pression as simple as possible and with little con-
`sideration for compression complexity, because
`compression can be performed by one of the net-
`work servers.
`One key consideration is how long the portable
`can function between battery rechargings. Ideally,
`it should be able to operate for one work day, or eight
`to ten hours ofbattery life. Given that conven-
`tional batteries typically possess 20 watt-hours for
`each pound of battery weight, and a limit of one
`poundof batteries in the portable, the entire portable
`can consume no more than 2 watts of power. Fur-
`thermore, projections of progress in battery tech-
`nologies show that only a 20 percent improvement
`in battery capacity will occur over the next ten
`years. Thus, power minimization becomes a seri-
`ous concern.
`
`
`
`2 For example, the EECS
`graduate student research
`facility at the University of
`California, Berkeley.
`
`Petitioners - Exhibit 1010 Page4
`
`The largest power consumer in current portables
`is the backlighting of flat-panel displays. As dis
`play technologies improve screen contrast, how-
`ever, this requirement will be significantly relaxed,
`implying that low-power techniques for implementing
`the analog and digital core circuitry are needed.
`With present-day technology — single-chip
`packaging using printed circuit boards for intercon-
`nection — one-third or more of the total power is
`consumed by the chip’s input/output (I/O) because
`P icoccllular networking ameliorates several im-
`the capacitances at the chip boundaries are usu-
`portant issues in providing portable multimedia-
`ally much larger than the capacitances inside the
`based communications systems. Many challenges
`chip. Typical values range from a few 10s of femto-
`remain, however, in building the required func-
`3 256 kb/s is typical ofthe
`farads at internal chip nodes, up to 105 of pico-
`tionality into terminal hardware.
`peak data rate aflorded
`farads at the chip interface attributed to pad
`A signal flow diagram for a terminal is shown in
`by most wireless LAN
`capacitances and board traces. To reduce the power
`Fig. 4. Incoming data can be one of three types:
`systems, and is reflective
`consumed in the I/O, low-capacitance, high—den—
`digitalvideo, screen graphics, or sampled audio. Out-
`of data rates used in exist-
`sity interconnect methods will be employed, such
`going data can be either voice or pen input. The asym—
`ing X-Ienninals; this value
`as the emerging multichip module (MCM) tech-
`metry in the data rates of the uplink and the downlink
`may be considerably lower
`nologies. MCM integrates many individual chip
`is clear; no high-rate signals are intended for trans-
`depending on the level of
`die into a single structure, reducing the size of
`mission from the mobile to the base station. The
`user activity.
`interchip capacitances to the same order-of—mag-
`hardware design must also reflect this asymmetry:
`
`
`____—_______r___—————————————
`67
`
`Implementing Portable Terminals
`
`IEEE Communications Magazine - December 1992
`
`Petitioners - Exhibit 1010 Page4
`
`
`
`DCT Algorithm
`
`Multiplies
`
`Additions
`
`
`
`each user, and hence reducing the bandwidth con-
`sumed by the overall system. Because high-reso-
`lution, full-motion Video is to be supported, trans-
`mission of 640-by-480 pixel images, digitized at 24
`bits/pixel, would require a bandwidth of 220 Mb/s
`in an uncompressed format at 30 frames per second.
`Thus, it is clear that video compression tech—
`niques are crucial in making wireless video trans-
`mission feasible.
`
`
`
`
`
`
`
`
`
`The video module performs the decoding and dis-
`play interface functions and converts a com-
`pressed data stream to an uncompressed video stream
`that is displayed on the LCD display. The decom—
`pression module can be implemented using a vari-
`ety of algorithms, such as transform-based schemes,
`vector quantization and subband coding. The
`selection of the algorithm for the portable termi-
`nal depends not only on the traditional criteria of
`achievable compression ratio and the quality ofrecon-
`structed images, but also on computational complexity
`(and hence power) and robustness to higher bit error
`rates. The choice of an algorithm to implement
`the decompression function is the most important
`in meeting the power constraints. The basic com-
`plexity of the computation must be optimized and,
`as shown in the next section, the ability to paral-
`lelize an algorithm will be critical.
`Most current compression standards (for exam—
`ple, JPEG and MPEG) are based upon the Dis-
`crete Cosine Transform (DCT). The basic idea in in-
`traframe schemes such as JPEG is to apply a two-di-
`mensional DCT on a blocked image (typically
`eight pixels by eight pixels) followed by quantization
`to remove correlations within a given frame. In
`the transform domain, most of the image energy
`is packed into only a few of the resulting coefficients,
`and compression is achieved by transmitting only
`a carefully chosen subset of the coefficientsWhilc
`these standards specify the use of DCT as the
`transformation to be applied, they do not specify
`the algorithm to be used. Table 1 shows a com-
`parison of a few algorithms that can be used to imple-
`ment the DCT [7—8]. Minimizing the operation count
`is important in minimizing the switching events
`and the hence the power consumption.
`A primary characteristic of the DCT is the sym-
`metric nature of the computation; that is, the coder
`and decoder have equal computational complexi-
`ty. However, an alternative compression scheme
`is that of vector quantization (VQ) coding, which
`is asymmetrical in nature and has been unpopular
`due to its complex coder requirements. The basic
`idea behind VQ coding is to group the image data
`into avcctor and quantize it. Fig. 5 shows a block dia—
`gram of a V0 coder. On the coder side, the input
`is first blocked into a vector (for example, a four-
`by—fourblock ofvideo is 16bytcs).This vector is com-
`pared to the entries in the codebook with the goal
`of minimizing the expected error or distortion
`between the input vector and its reproduction for
`a given bit rate. The codeword corresponding to
`the closest match (or the match that minimizes the
`distortion) is transmitted (in this case, the index is
`one byte long and 16:1 compression is achieved). On
`the decoder side, a simple memory look-up is
`used to reproduce the image data.The design of the
`codebook and distortion measure have been exten-
`A 5 stated above, frequency reuse is only one
`means of reducing spectrum consumption. Con-
`sively discussed in literature. Clearly, the distor—
`comitant with frequency reuse is the idea of re—
`tion computation is much more computationally
`intensive than the decoder, as the entire decoder
`ducing the amount of physical bandwidth needed by
`
`
`
`
`OutpUt
`
`m _ Distortion
`
`
`calculation
`|00l<UP
`(4 x 4)
`(4 X 4) ,
`t _
`
`‘
`,
`,
`codebook
`
`(2 56 levels)
`
`
`
`,
`
`,
`
`»
`
`Codebook
`
`(256 levels)
`
`
`
`I Figure 5. Block diagram ofa VQ coder.
`
`nitude as on-chip capacitances and minimizing the
`power consumed in the I/O drivers. Thus, with MCM,
`the majority of the power is consumed within the
`functional core of the chip itself, as opposed to
`the interface. Also, because the packing density
`has increased, and with the ever—decreasing size of
`CMOS circuitry (down to 0.2 um line widths), over
`1010 transistors can be placed within a single
`eight-by—eleven-inch MCM substrate. Area con-
`straints imposed by available silicon are no longerof
`great issue, allowing for greater possibilities in power
`optimization, as we will discuss later.
`For analog RF transceivers, however, there are
`other design considerations beyond low-power imple-
`mentation. Due to size considerations, traditional
`discrete element design is not feasible for a small,
`portable unit such as the proposed multimedia
`terminal; single-chip integration techniques that
`exploit advances in silicon CMOS (asopposed to gal-
`lium arsenide, GaAs) must be explored, to address
`cost and manufacturability concerns. Likewise,
`the fact that digital circuitry is readily available on-
`chip also opens up new possibilities: analog per-
`formance requirements can be reduced at the expense
`of increased digital signal processing.
`To examine these implementation details
`more fully, we will discuss three distinct design issues.
`The first is image compression, both as a means
`of spectrum reduction and as an example of how
`the choice of algorithms can take advantage of
`the aforementioned asymmetry to reduce power con—
`sumption. Second, we will show how low-power
`digital system design leads to large reductions in
`powerconsumption. Third, we will analyze the design
`of analog RF transceiver to exploit monolithic
`integration techniques and the underlying digital
`nature of the transmitted signal.
`
`Image Compression
`
`68
`
`IEEF. Communications Magazine ' December 1992
`
`Petitioners - Exhibit 1010 PageS
`
`Petitioners - Exhibit 1010 Page5
`
`
`
`Algorithm
`
`Full—search”
`
`Coder per pixel
`
`Decoder per 16 pixels
`
`ubtract.‘ multiply Lad'd”
`emery access
`
`:11 memoryacc
`
`1 memory access
`
`
`
`is nothing more than a lookup table; however,
`algorithms have been proposed that reduce the com-
`putational complexity in the coder with little loss
`in reconstructed image quality [9]. Table 2 shows the
`requirements for two strategies.
`The simple memory-lookup decoder of V0 algo-
`rithms makes them well-suited for single-encoder,
`multiple—decoder systems (in which video is com—
`pressed and stored once, but is accessed multiple
`times by multiple users). If one—way video commu-
`nication is desired, the V0 solution provides a means
`of implementing real—time decompression requir-
`ing little computation and power. For portable ter-
`minals, supporting one-way communication in which
`a user accesses various databases on the wired net-
`work is a very important feature.
`The algorithms described above are based on
`removing redundancy inside a frame (intraframe),
`and are quite robust against higher bit-error rates.
`For a typical channel BER of 10"", this translates to
`an average of only a couple of local regions per sec—
`ond having corrupted image data.This magnitude
`of error is acceptable for video as human perception
`is not sensitive to such small, local errors. Like-
`wise. the errors do not accumulate, because the next
`frame will be fully transmittedwith its own local error
`independent of the previous frame. On the other
`hand, interframe schemes remove temporal
`redundancy (i.e., correlation between frames)
`and are not as robust against bit errors. The basic
`idea in these schemes is to apply DCT or VQ on
`the difference of the current image and the previ—
`ous image. Since only the difference information
`is transmitted, a bit error can cause many regions
`to be corrupted. Also, errors accumulate in the
`case of interframe coding, making it less desir»
`able; by “resetting” the errors every so often by trans—
`mitting a complete frame, however, such differential
`error propagation can be alleviated. However, it is clear
`that the “best” algorithm for a wireless environ-
`ment will likely be quite different than one cho-
`sen for the low-BER situation assumed by the present
`compression standards.
`
`Tree search with a
`differential code-book
`
`8 multiply, add
`8 memory access
`
`I Table 2. Computational complexity of VQ algorithms for a vector size of 16
`(4x4).
`8.0
`
`
`
`Normalizeddelay
`
`9":3
`
`«hb
`
`
`2.0
`
`I Figure 6. Plot of normalized delayys. supply volt-
`age at Vdd (delay at Vdd = 5 Vis normalized to I).
`
`power-delay product) is equal to Cavg-lfidz, where
`Can is the average capacitance being switched per
`clock and Vdd is the supply voltage. Average power
`consumption is given by the product of the ener-
`gy per transition and the frequency of operation, fdk.
`Power is also consumed by short circuit currents
`(which arise from a direct current path from sup—
`ply to ground during switching) and by subthresh-
`old leakage currents (parasitic currents that flow
`when the transistors are supposed to be turned
`off). The total power dissipation in CMOS cir-
`cuits is summarized in the following equation:
`
`Prom! = Cavg' I{ltlz'fclk + Isc' Vlid
`+ Ileakage ' Viid
`
`Eq' 1
`
`Low Power Systems Design
`massive amount ofcomputing resources willbe
`required to achieve the performance desired
`ofportablemultimedia terminals—each unit needs
`to support both complex modulation schemes and
`sophisticated data decompression algorithms. Like-
`wise, the power nceded to drive these functions can
`easily be prohibitive for portable operation. Total
`power consumption must be minimized and the
`required throughput of the overall system must
`be maintained. However, because processing is
`bounded by real-time constraints, as in image decom—
`pression and channel equalization, once the through—
`put performance is met there is no advantage is
`making computation any faster, opening up a
`major degree of freedom to the designer. This
`section comprises an overview of techniques for min-
`imizing the power consumption in digital CMOS cir-
`cuits.
`
`By careful circuit and technology design, the short
`circuit and leakage components can be mini-
`mized and the goal of low power design becomes the
`minimizing of Cavg, Kid, and fclk, while retaining
`the required functionality.
`The quadratic dependence of energy on volt-
`age has been experimentally verified for various
`circuits, and it is clear that operating at the lowest
`possible voltage is most desirable for optimizing the
`energy per computation. Unfortunately, reducing
`the supply voltage comes at the cost of a reduction
`in the computational throughput. This is seen
`from Fig. 6, a plot of normalized delay vs. Hid for
`atypical 2pm CMOS gate, with the gate delays
`increasing as the supply voltage is dropped. Even
`though the exact analysis of the delay is quite com-
`plex, it is found that a simple first-order deriva—
`tion adequately predicts the experimentally deter-
`mined dependence quite well and is given by:
`
`Sources of Power Dissipation
`CL”,
`szcLyzg=
`The main contribution to power consumption in CMOS
`KP(W /L)(Vdd ~V,)2
`I
`circuits is attributed to the charging and discharging of
`where CL is the load capacitance, I is the current
`parasitic capacitors during logical transitions. The
`average switching energy of a CMOS gate (or the
`drive, K1) is a process dependent parameter, W/L
`__________—__‘__’_————-———-——
`
`Eq. 2
`
`IEEE Communications Magazine ‘ December 1992
`
`69
`
`Petitioners - Exhibit 1010 Page6
`
`Petitioners - Exhibit 1010 Page6
`
`
`
`Architecture
`
`optimization
`
`will play a
`
`critical
`
`role in
`
`optimizing
`
`power
`
`dissipation
`
`by enabling
`
`low-voltage
`
`operation.
`
`
`
`is the size or the strength of a transistor, and V, is
`the threshold voltage (the gate-to-source voltage
`at which the transistor “starts” to conduct cur-
`rent) [16]. It is clear from the a