In the effort to facilitate and standardize the flow of data into the home, the delivery method and management of services are crucial considerations. In many cases, delivery depends on two-way communication and guaranteed control of proprietary media content.

This paper focuses on handling the multimedia content after it is delivered. The various data streams to consolidate and manage are ranging from voice and video over IP, cable/satellite TV with DVR functionality, and Internet services to on-demand media delivery and online gaming.

The wide variety of formats and the potential need to interchange them pose unique challenges for a multimedia center (MMC) or a multimedia home platform (MHP). In this paper, the term MMC designates the actual hardware (center) and the MM system (MMS) integrates the software component of the design.

Traditionally, each unique media function is handled by a dedicated device and often a dedicated communication line. The result is the redundant duplication of, for example, a digital cable delivering TV programming, an Internet provider with IP phone service, and a regular phone line all using essentially the same infrastructure but different delivery methods and different dedicated devices. The current trend is to consolidate this multiple functionality into a single device. In many instance a home computer plays this role, but in a consumer friendly way the challenges of price, simplicity, and efficiency remains open.

To address these challenges, this paper presents a general marketing case for an MMC implementation using software techniques. It considers PowerPC architecture as a candidate for such an implementation and provides performance measurements to support these claims.

Current implementations
As an example of the consolidation trend let us point to several recent offerings in the area. Several solutions built around the Intel Viiv Digital Right Management (DRM) [1] platform have been recently announced (PC Fusion 2 [2], The Ace® LH series of Digital Entertainment systems [3]; AVidius® Media Systems [4]). All of thse closely resemble a general computational platform with a SW layer to provide multimedia handling capabilities and a unified interface. The general purpose dual-core CPU (Core Duo) is used in this model.

Another example is Viiv's direct rival " AMD Live! [5] with a very similar approach. Both platforms illustrate the idea of an implementation based mainly on a software solution with substantial computation power dedicated from a general purpose core. At the same time both platforms incorporate Digital Right Management and will be using Microsoft Windows OS (for instance Windows XP Media Center Edition " XP MCE).

Earlier system examples include the following [21]:

1) THOMPSON IP900 (introduced back in 2003), is a MIPS-based SoC, while the next-generation IP100x is based on Intel x86. Both systems are nominally IP STB boxes, but the IP1000 decodes MPEG-4 and H.264 and the IP900 handles only the MPEG-2 stream. (See Table , below) [21]

2) SAMSUNG SMT-7000S is based on Intel Celeron and is marketed by Samsung as an IP/DSL STB.

3) FUJITSU SIEMENS Series 200, 300 IP-STB (uses a 1.25 GHz Intel CPU). Another recent example - the ACTIVY Media Center 500 series [6] (shown in Table 1 below)..

4) Handan (Korea) uses the Celeron 733 MHz or 1 GHz CPU.
5) Humax (Korea) uses the VIA Eden CPU (800MHz) [7]
6) Softier [8] IP STB based on TI DSP and Linux OS.

Table 1. Thompson IP900/IP100x

All of these examples illustrate an approach that uses a general-purpose CPU with limited peripherals to implement MMC.

On the other end of the spectrum are numerous systems with an extremely light-weight general-purpose core to execute control functions and several peripherals serving as a dedicated hardware solution. These systems represent today's vision of a collection of many dedicated components, each executing a specific function. These platforms target the high volume, low price market and tend to be very rigid in their implementation.

Feature set
There is a wide design space for MM systems, as well as general disagreement about their definition and classification. The design space can be loosely defined as several tiers according to implementation, cost, and feature set, as shown in Classification of MM Systems by Functionality 1. The DRM (Digital Right Management) is a part of any system by default.

Figure 1. Classification of MM Systems by Functionality

To define an exemplary tier 3 MMS, which will be considered in this work, we describe the desired feature set in greater detail:

1) Set top box (STB) with digital video recording (DVR) functionality. Ability to receive (terrestrial, satellite, IP TV carriers), view, pause, rewind and record live TV (timeshift). The system handles two SDTV streams simultaneously, one viewed (deinterlacing, if needed) and another recorded or overlaid as a picture-in-picture. There is programming guide retrieval with advanced DVR management: schedule recordings ahead of time, following programming changes, and resolving conflicts between multiple recordings.

2) DVD, CD, VCD playback, management (transcoding) and writing/burning.

3) Media cataloging and reliable storage that can seamlessly interface to NAS or implement an on-board RAID system. Redundancy enabled storage such as RAID can be a significant marketing point, but is also very sensitive to the form-factor of the system. For the current design, simple high-capacity HD is used for content storage. RAID is targeted as future work.

4) Internet access console that includes browsing, automated information retrieval (weather, news headlines) and elements of data management (headline analysis for key words of interest, offensive content filtering). Functionality may include networking appliance capabilities: in-house router functionality, software firewall maintenance, and VPN support. The proposed system handles only automated data retrieval, which depends on external content providers such as the CNN web news site and the weather channel web site.

5) On-demand delivery of media such as movies, music, and games. The delivery requires digital rights management. This feature requires external provider interaction, so it cannot be tested and therefore is not implemented in our design.

6) Gaming support in both the on-demand and game console modes. This option may demand major changes in system design, thus defining many other aspects of the system such as the level of graphic card support. We use the system only as a game console and implement several open source games.

7) Voice over IP and Video over IP functionality such as video conferencing or simply digital phone service using an IP connection. This functionality puts a specific demand on a variety of protocol support, both for connection management and for voice/video codec (SIP, H263, G711, G729, and so on). This feature is fully implemented and allows us to conduct a full duplex video call.

8) A primitive artificial intelligence (AI) functionality is a bonus to improve the human interface and feel. It can assist in searching a program guide based on close caption or programming guide descriptions. It enables easy, hands-free voice recognition and text-to-speech, navigation among features, and so on. The text-to-speech feature is implemented but not currently integrated with the rest of the system. The AI component is also reserved as future work.

Most of these features require high computational power and are well suited to a powerful CPU plus software solution. The form-factor for such a system, in addition to likely in-room positioning, would yield the benefits of a noiseless (fan-less) low power system. Given these requirements, the ratio between processing power and power dissipation as well as cost sensitivity is extremely important.

Although many of these functions overlap with the role PCs play today, in MMC they are limited mainly to consumption, not creation, of the media, non-modular design, and emphasis on a seamless human interface.

A variety of open source projects implement MM server functionality. Most are collections of multiple features as discussed in Section 2, but two of them deserve inspection: MythTV [9] and Freevo [10]. Both projects implement a digital video recorder (DVR) for an STB-like design and both have certain advantages and drawbacks.

After both applications were tested in two independent implementations, MythTV was chosen for its final design. A significant advantage of the MythTV project is its configurability. Additional plugins/components can be integrated seamlessly into the common GUI front end, which in turn is easily customized to meet vendor requirements.

This system yields the benefit not only of design flexibility but also easy distribution between the main server and multiple satellite stations. The main server runs the MythBackend with MySql database and any high intensity jobs, while the satellite station runs MythFrontend thin client, which receives only processed content and outputs it to a user console (regular TV or progressive monitor).

Thus, the actual compute intensive processing can be physically partitioned from the access console, which can be very important in a design space where too much processing power and too many features are concentrated in a single location (tier 4 in our classification " See Figure 1, above). In this scenario, a backbone system with RAID storage and a powerful processing core can reside outside the entertainment room in a basement or a closet with light profile satellite stations located throughout the house.

It is worth mentioning that open source multimedia server projects are sometimes criticized for their installation complexity and high initial investment in the software configuration. These arguments are voiced by Linux enthusiasts who configure the software for their custom systems. In a production environment, these criticisms do not apply. However, updates and maintenance of previously released software components requires attention to the version of the underlying Linux OS.

Multiple vendors package free Linux code and redistribute it under various brand names and different packaging options [11][12][13]. Such packages include preselected (and often precompiled) sets of features/applications.

One notable exception is Gentoo Linux [14], which markets not a distribution but rather a system knowledge tool that allows designers to create their own custom distributions tailored to a specific platform and need. The designer is in charge of selecting the required components, and the system knowledge tool guarantees that all the appropriate packages are delivered and all dependencies are respected.

The high-level end application (such as MythTV) check out prompts an automated yet very complex process of verifying components and package presence on the system, their interdependency, necessary source code delivery, and a custom build followed by installation of required components—all from a single command.

Not only is development time minimized but also there is a guarantee that any future updates and/or bug fixes in any of the components prompts updating and rebuilding as needed. The result is an improved version of the final product. With an MMS built with a general-purpose CPU at a known location, adding new features and fixing problems becomes very similar to the software update mechanism used by most commercial software vendors. For details, refer to [15]

Platform Specific Implementation
In selecting the PowerPC architecture to implement the multimedia server, we note that the main advantage of the PowerPC G4/G4plus class cores [16] is the excellent performance-to-power ratio. The PowerPC architecture system can use passive cooling to provide a zero-noise, fan-less implementation that meets the aggressive goal of high processing power to handle multiple HDTV decoding streams in software or conduct a video conference while recording an HDTV stream.

The PowerPC architecture is inherently friendly to multimedia content. All G4/G4plus class processors [16] offer the advantage of the AltiVec SIMD unit [15][17]. AltiVec technology was originally conceived and developed by a design consortium consisting of Apple Inc., Motorola SPS (now Freescale Semiconductor Inc.), and IBM Inc. to accelerate multimedia processing.

The AltiVec technology augments the processing power of a sophisticated superscalar core by offering additional vector/parallel programming capabilities that can significantly speed-up many classes of communications and multimedia applications: IP telephone gateways, multi-channel modems, VPN systems, speech processing systems, echo cancellers, digital image and video processing systems, scientific array processing systems, and so on.

The 128-bit wide AltiVec engine remains the most sophisticated and powerful option for SIMD extension processing to a general-purpose superscalar core. Numerous applications are written for it, and most trade-offs and design nuances are well understood. All new G4/G4plus PowerPC cores from Freescale include AltiVec technology, which is fully backward-compatible with all the software developed for it [16].

To illustrate the potential of the AltiVec unit, we reference the independent EEMBC benchmarking consortium [18], which has published multiple benchmarking scores in many application categories over the years. Of interest for the work discussed here is the Digital Entertainment subcommittee, which contains standardized benchmarks for MPEG2 and MPEG4 encoding end decoding, MP3 decoding, JPEG processing and much more [19].

This subcommittee has produced a total of 15 benchmarks with 69 data sets grouped for convenience in several so called sub-marks: MPEG Decode Mark, MPEG Encode Mark, Encryption Mark and Static Image Mark. For details, consult [19], [20]. Additionally, there is a single aggregate score for the entire 69 benchmark/data set permutations, known as Digital Entertainment Mark or DENMark.

Figure 2. Breakdown of the EEMBC

Here we present scores for several comparable general-purpose CPUs, including IBM PowerPC 750GX and 970FX cores (970 FX runs in 32-bit mode), AMD Geode NX1500, Analog Devices BF533, and two CPUs from Freescale Inc. (See Figure 3, above and Figure 3, below).

Figure 3. Aggregate Digital Digital entertainment sub-marks

The MPEG_Decode_Mark is composed of scores for MPEG2, MPEG4, and MP3 decoding. The MPEG Encode mark comprised of MPEG2 and MPEG4 encoding scores. Static ImageMark is a collection of Compess/Decompress JPEG image and several color processing filters. EncryptionMark is of little interest for our current discussion. The legend for tested parts is as follows: vendor name_part name_core frequency_memory frequency. In addition, Freescale processors have two score variations:

1) Regular. Known in EEMBC as out-of-the-box (OOB) and used by default for all certifications. It is obtained by a simple compilation of the standard C code with a fixed set of compiler options.
2) Optimized (OPT). Adds core-specific intrinsics to the C code.

The difference between OOB and OPT is what we want to concentrate on " it shows the performance advantage of AltiVec technology and yields a 217 percent average performance increase for the entire benchmarking suite. Figure 3 also illustrates the single aggregate for all the scores " the DENmark which exhibits a similar trend. For details and an individual benchmark breakdown, refer to the EEMBC web site [18][19].

For the Freescale products, typical power dissipation while running this code measures at about 20 watts for MPC7447a at 1.4 GHz, and about 15 Watts (L Spec) for MPC7448 at 1.7GHz [16]. This low power dissipation is well within the limits for a passively-cooled simple heat sink solution. The AltiVec technology measures at a negligible 5 percent of the total power dissipation while delivering a several times performance boost.

This amalgamation of a high performance superscalar CPU with the 128 bit wide parallel SIMD engine, along with low power dissipation, makes the general-purpose PowerPC core very attractive for multimedia server implementations.

Most algorithms in our system use AltiVec technology in different degrees, which enables the design and permits the use of a much slower CPU for the function.

Using SIMD to handle multimedia content
The most likely, and one of the most demanding, algorithms to be encountered by the MMS is software MPEG decoding for broadcast TV and DVD playback. The performance numbers based on EEMBC MPEG2 are representative of a DVD playback in both resolution and decoder configuration. MPEG decoding for TV broadcasting is very similar.

Table 2. ATSC DTV standards

Generally, MPEG2 is most commonly used as the compression algorithm for both SDTV and HDTV broadcasts. Although MPEG2 supports up to 4:2:2 YUV chroma subsampling and 10-bit quantization, which EEMBC uses in most data sets, DTV broadcasts use only 4:2:0 and 8-bit quantization to save bandwidth, which in turn changes CPU workload slightly. Some HDTV broadcasters plan also to use MPEG4, which offers a slightly higher compression rate for a given perceived quality and/or possibly multiple channels to deliver a high-end HDTV signal.

The North American Advanced Television Standard Committee (ATSC) defines 18 formats for DTV (Table 2, above). Here the MPEG2 data rate is a close approximation and can vary slightly. Also, notice that the actual NTSC signal has some overhead for audio and transport organization so the end data rate varies to some extent. HDTV is using the Dolby Digital (AC-3) format to support 5.1 surround sound, theater quality audio track. Nevertheless the data rate for audio track is negligible compared to the rate of the video signal.

Figure 4. Pegasos ODW Block Diagram

MPEG2 MP ML (Main Profile, Main Level) decoding in software can be roughly separated into four operations, some of which are parallelizable. That is, they are locally data independent and can be executed in parallel (Figure 4, above). As the EEMBC scores in Table 3 below, AltiVec technology can be used very effectively to perform these decoding operations. This optimization can dramatically accelerate the performance of individual algorithms. For example, AltiVec unit can accelerate 2D IDCT in a stand-alone mode by nearly 500 percent (five times).

Table 3. MPEG2 Decode profile

Dequantization (iquant) can be accelerated by about 200 percent, and motion estimation can be improved by 200 percent. Nevertheless, when all components are integrated for TV MPEG2 decoding, the overall speedup (an average of 10 data sets, different from EEMBC) is about 245 percent [22].

This increase is in line with the performance gains reported by the EEMBC set of benchmarks for DVD-like MPEG decoding. MPEG2 encoding is much more work intensive than decoding and presents more opportunities for vectorization with AltiVec technology. For encoding, the average speedup is measured at around 480 percent [22]. Encoding is a factor in MM system when transcoding is performed between various formats. In a similar analysis for digital audio decompression (playback) tested for an MP3 algorithm with DVD quality, AltiVec unit delivers an average 220 percent speedup compared to the original implementation [22].

All these examples underscore the effectiveness of AltiVec technology in different multimedia formats processing and the overall suitability of the PowerPC architecture for an integrated home media center.

System-level design considerations
In a reference design with real-time performance analysis and workload balance trade-offs, the bulk data flow dictates all design trade-offs. To this point, we have considered a general-purpose CPU with a software-oriented implementation as more flexible and application independent. Benchmarking measurements justify the use of PowerPC core with AltiVec technology in the design. To address the cost/feature trade-off and conduct system-level benchmarking, we proceed to implement the system.

This system shown below in Figure 4 above is an off-the-shelf design for a general-purpose Linux desktop PC that is not specifically targeted for multimedia applications, so some trade-offs are not addressed. However, as a proof of concept, this system serves its purpose. The system is designed by Genesi Inc. [23] and is in fact the Pegasos Open Desktop Workstation (ODW) [24] platform, meaning that all design and implementation details (gerber files, schematics, component list, and so on) are freely available [25] and, with the open source software layer implementation, add up to a truly open design.

In this design exist some functional latitude to configure applications to explore potential design space and illuminate important design trade-offs. One example is the use of the CPU for the graphical rendering, minimizing the requirement for video card functionality. Indeed, the multimedia system mainly addresses 2D graphical applications, with limited support for gaming applications and simple 3D graphics (based on OpenGL). Therefore, we can use the simplest video card in a frame buffer (FB) mode (just capable of displaying memory buffer content without processing it).

One of the most important trade-offs in such a system design is where the MPEG2 TV signal decoding takes place. To guarantee that the TV signal is distributed freely throughout the system built around PCI bus (PCI 32 bits at 33 MHz bandwidth = 133 Mbps = 1Gbps), the raw SDTV (data rate is about 105 Mbps; the raw HDTV is up to 1.4 Gbps) is delivered by the TV tuner in compressed form.

The TV tuner used in this case is Hauppauge WinTV PVR 500MCE. The compressed data rate is about 2 Mbps for SDTV and up to 18 Mbps for HDTV (See Table 2). Of course, the compressed data must be decompressed/decoded prior to rendering, a function frequently outsourced to a coprocessing unit. In addition to decompressing, the following steps may be necessary:

1) Deinterlacing to display the data on a progressive screen, a time-consuming algorithm that can consume a non-trivial amount of processing power.
2) Adding overlay information.
3) Blending with another stream for a picture-in-picture. Please note, that the picture-in-picture feature requires two TV tuners, which are available at the Hauppauge WinTV PVR 500MCE, and simultaneous decoding of two MPEG2 streams.

One option available in our system is to take the MPEG2 encoder output from the tuner card, use the DMA controller to transfer the data directly to the ATI Radeon 9250 built-in MPEG decoder, and output the decoded video in interlaced form for display. The use of DMA minimizes the amount of processing required from the CPU, freeing the CPU to perform other tasks.

The disadvantage of this approach is the interlaced output unless additional hardware is used, difficult overlay and picture-in-picture construction, decoding of only one stream at a time, and the actual need to use the Radeon 9250 or another video card with an MPEG decoder in the final design. The preferred option is to allow the PowerPC core to do all the work and use the Radeon 9250 in frame buffer mode only (to be potentially replaced by a cheaper video card).

Contrary to the situation with the graphic card, the Creative Sound Blaster Audigy2ZS (SB 0350) audio card is used to its full potential. It delivers a true 5.1 channel surround sound experience for appropriate media. The software sound manager is ALSA [26].

DVD playback algorithmically is similar to TV decoding. The player for our system is Xine [27], which is built using the FFMPEG project [28]. For CD and VCD as well as most other video file formats (MPEG4 derivatives), playback does not represent a significant design challenge and can be handled by the same codec.

Transcoding on the other hand might require non-standard output formats and/or resolutions, but due to the nature of the transcoding process, it does not have to be done in real time, so critical performance requirements are relaxed in this case.

Another interesting aspect of this design is the seamless integration of the video, voice over IP, and instant messaging into the system. MythPhone [29] is the default plugin used in MythTV to provide this functionality (see Figure 6). Its feature set includes the SIP communication protocol, G.711 or GSM for voice, and H.263 for video calls.

Another option is the Gnomeeting project [30] that is based on the OpenH323 Library [31] and supports numerous formats, including H261 (video), G711, G726, G723.1, GSM (audio), and a variety of connection protocols. Also, Gnomeeting is fully interoperable with Microsoft NetMeeting.

Results
For experimental part of the work we have conducted study of several common scenarios which are summarized in Table 4, below. The purpose is mainly a system level benchmarking " making a realistic estimate of the system capabilities in addition to existing core-centric benchmarks. The TV signal in this case is US Cable NTSC signal with 480i resolution, which in most cases is getting converted to 480p (deinterlacing).

Table 4. CPU Workload results for common system level scenarios

As a result of this experiment, the system was proven capable of recording one SDTV channel (no rendering required, just write to a file, no deinterlacing) and playback of another channel at the same time. This amount of processing is representative of the peak workload for the system (for the given tier three category) while handling TV programming.

Also an indirect comparison between scenario five myth frontend workload (no deinterlacing done) and any other scenario where the deinterlacing is performed demonstrates the non trivial amount of work to interface a progressive monitor. This might be worth another design tradeoff consideration whether to address a standard interlaced TV or high definition progressive monitor.

The amount of workload required for MPEG2 decode is just slightly different between the TV and DVD playback (scenario four vs. any other TV playback) and hovers in a single digit percentage difference. Actual rendering adds surprisingly little to overall workload (15-18% for the whole X windows functionality, which includes rendering) and further supports the claim about PPC capability to absorb this function.

Conclusion
In this work we attempted to present an appealing case for defining a multimedia consumer-oriented system and justifying PowerPC core with AltiVec technology as an appropriate platform for its implementation. The system was built using open source software and off-the-shelf components to serve as a test case or a general-purpose prototype for system level trade-off analysis. The results demonstrated that a 1GHz PowerPC core with AltiVec technology is very capable of handling two SDTV data streams simultaneously, with CPU cycles left to spare. Future work will implement HDTV functionality on the same or a similar system, with RAID integration and human interface improvements.

Sergei Y. Larin is senior applications engineer at Freescale Semiconductor and Pieter Van den Abeele is affiliated with Genesi Inc.

This article is excerpted from a paper of the same name presented at the Embedded Systems Conference Silicon Valley 2006. Used with permission of the Embedded Systems Conference. For more information, please visit www.embedded.com/esc/sv.

References
[1] Intel Corp. www.intel.com
[2] Digital Home PC Inc. - www.digitalhomepc.com/fusion.htm
[3] Ace Computers -www.acedigitalhome.com
[4] Computer Connections - www.paconnect.com/avidius.asp
[5] Advanced Micro Devices, Inc. www.amd.com.
[6] Fujitsu Siemens Computers
[7] Via Technologies, Inc. - www.via.com.tw/en/products/processors/eden-esp/
[8] Softier, Inc. - www.softier.com/iptv_stb.asp
[9] MythTV Project " www.mythtv.org
[10] Freevo Project - freevo.sourceforge.net
[11] Red Hat, Inc. - www.redhat.com
[12] Software in the Public Interest, Inc. - www.debian.org
[13] Terra Soft Solutions, Inc. - www.yellowdoglinux.com, www.terrasoftsolutions.com
[14] Gentoo Foundation, Inc. " www.gentoo.org
[15] Sergei Larin " "Home Media Center Built with Linux on PowerPC Architecture" Embedded Systems Conference and Design Seminars 2006, April 3-7, 2006, San Jose McEnery Convention Center, Ca.
[16] Freescale Semiconductor, Inc. " www.freescale.com/powerpc
[17] Freescale Semiconductor, Inc. " www.freescale.com/altivec
[18] Embedded Microprocessor Benchmark Consortium - www.eembc.org
[19] Embedded Microprocessor Benchmark Consortium, Digital Entertainment subcommittee
[20] Sergei Larin " "EEMBC V2 Benchmarking Overview "- (SBG844) " Freescale Technology Forum (FTF) 2005 - Orlando, FL
[21] In-Stat Marketing Research Report - www.instat.com
[22] Sergei Larin " "AltiVec in Multimedia Applications "- (H1125) " Smart Networks Developer Forum
(SNDF) 2004 - Dallas, TX
[23] Genesi USA Inc. - www.genesippc.com
[24] Open Desktop Workstation -www.pegasosppc.com/odw.php
[25] Power.org Consortium - power.org
[26] The Advanced Linux Sound Architecture Project
[27] The Xine Project " www.xinehq.de
[28] The FFmpeg Project - ffmpeg.sourceforge.net
[29] Paul Volkaerts " MythPhone Project " www.zen13655.zen.co.uk/mythphone.html
[30] GnomeMeeting Project " Ekiga - www.ekiga.org; www.gnomemeeting.org
[31] The OpenH323 Project " www.openh323.org