AMD Fusion Family of APUs: ™

Enabling a Superior, Immersive

PC Experience
Abstract:
The AMD Fusion™ Family of Accelerated Processing Units (APUs) is coming. These
Table of Contents
processors’ compatibility with Microsoft® Windows® 7 and DirectX® 11 is designed
Introduction . . . . . . . . . . . . . . . . . . . . . 2 to ensure that they will provide an outstanding out-of-the-box experience for
What is an Accelerated those who use PCs built on them. Working in conjunction with the advanced
Processing Unit?. . . . . . . . . . . . . . . . 3 x86 processor cores, an APU’s multiple vector cores enable software developers
to create innovative new applications that expand usage models and make
Why all the fuss about
“vector processing”?. . . . . . . . . . . . 4
PCs fast, easy to use, and more intuitive. The recent introduction of new tools
(DirectCompute and OpenCL™) for thread-level and data-parallel applications
The ABCs of vector development simplifies the task of creating these applications. The development
processing. . . . . . . . . . . . . . . . . . . . . . . 5
platforms and tools are scheduled to be widely available. The rest is up to
How hard is it to accelerate entrepreneurial developers and PC manufacturers.
software using GPU
computing resources? . . . . . . . . . . 6

Does this stuff really work? . . . . . 7

What can APUs do for

traditional workloads?. . . . . . . . . . . 7

Summary. . . . . . . . . . . . . . . . . . . . . . . 8

Nathan Brookwood, Research Fellow, March 2010

Insight 64 thanks AMD for its financial and technical assistance in the creation of this white paper.

AMD White Paper: AMD Fusion™ Family of APUs

 Introduction extra transistors have been used to increase the size of
Imagine a PC that: on-chip caches and add more x86 processor cores to
»» Recognizes your face to login designs, making today’s CPUs the fastest processors
»» Recognizes your gestures without a remote ever. Even the slowest contemporary CPUs have more
»» Responds to your touch or voice to do your bidding than enough performance to handle traditional office
»»Supports bi-directional hi-definition video chat over productivity, Internet browsing and e-mail applications,
links with limited bandwidth which long ago ceased to be limited by CPU speed. But
»»Finds and tags the photos and videos in your library as fast as they are, today’s CPUs lack the performance
that contain particular faces, places or objects to deliver a vivid, modern computing experience on their
»»Helps you sort through your photo libraries to eliminate own. The latest applications require CPUs that can deal
duplicates saved with different file names with vast amounts of data and require hundreds, if not
»»Enhances the videos you’ve created with regard to thousands of individual threads to manipulate the
color, focus and image stability massive databases needed to recognize an object in
»»Takes you to the right page in IMDB or Wikipedia a scene, the meaning in a sentence, or an anomaly in
when you point to an actor in a movie or a figure in an x-ray image. Not surprisingly, traditional CPU
a news program architectures and application programming tools
»»Up-scales even low-quality content to seamlessly optimized for scalar data structures and serial algorithms
match the capabilities of your HD display fit poorly with these new vector-oriented, multi-threaded
»»Adds stereoscopic 3D realism to 2D content data-parallel models.
»»Supports immersive, multi-monitor 3D gaming
experiences Fortunately, innovative architectures and tools better
»»Sells at price points well within reach of the suited for these new workloads have emerged. Graphics
mainstream consumer. processing units (GPUs), originally intended to enhance
3D visualization, have evolved into powerful,
Many of these capabilities exist today piecemeal in programmable vector processors that can accelerate
labs, running on expensive, workstation-class computers a wide variety of software applications. Software tools
that cost as much as tens of thousands of dollars. Why like DirectCompute and OpenCL permit developers to
haven’t we progressed further, faster in delivering these create standards-based applications that combine the
capabilities to the mainstream? The semiconductor power of CPU cores and programmable GPU cores, and
industry prides itself on rapid improvements in system run on a wide variety of hardware platforms. A few
performance, but hardware that runs fast enough to ambitious independent software vendors (ISVs) have
enable these advanced capabilities still costs far too already added support for these new vector capabilities
much to enable high-volume deployment. Software into their most advanced products, even if they had to
developers, always tuned to market realities as well as structure their code around proprietary hardware and
technology, have focused their efforts on applications software interfaces to get the job done.
that run well on the dual- and quad-core x86 processors
that comprise the bulk of today’s mainstream system Advanced Micro Devices’ (AMD’s) forthcoming
offerings. But change is in the air; in 2011, affordable Accelerated Processing Units (APUs) build upon this
mainstream systems that can support these advanced momentum and take PC computing to the next level.
capabilities are set to enter the market. These new processors are being designed to accelerate
multimedia and vector processing applications,
You’ve probably heard this story before. Every two years, enhance the end-user’s PC experience, reduce power
advances in semiconductor technology allow chip consumption, and offer a superior visual graphics
architects to double the number of transistors they can experience at mainstream system price points. Insight 64
fit in a given area of silicon. Over the past decade, these would expect no less. More importantly, these APUs will
1 Subject to restrictions and limitations; fees and charges may apply.
See http://www.dell.com/us.en.business/servers/pedge_t105/pd/aspx?refid=pedge_t105&cs=04&s=bsd.
2 Additional hardware
2 or software may be required for remote access. Internet access required. AMD White Paper: AMD Fusion™ Family of APUs
 enable ISVs to create new generations of applications notebooks, and netbooks and slates. Both of these APUs
and user interfaces limited perhaps only by the will combine multiple superscalar x86 processor cores
inventiveness of their developers, rather than by the with an array of programmable SIMD engines leveraged
constraints of the traditional CPU architectures that from AMD’s discrete graphics portfolio.
have dominated the computer industry for decades.
Figure 1. To enable the immersive PC usage models that users
demand requires both CPU and GPU collaborative computing
In this white paper, Nathan Brookwood of Insight 64
explores the world of these new APUs for AMD. How Multi-Core With External
do they differ from today’s CPUs and GPUs? Which
CPUs APU GPU Sub-Systems

applications benefit the most from this technology? CPU GPU

Can mere mortals harness their power? Even more SERIAL VIDEO
importantly, what steps should PC manufacturers and GENERAL DATA
ISVs take today to ensure they will be ready for the new PROGRAM PROCESSING
wave of accelerated processing units when it begins TASK
to crest, currently expected sometime in 2011? DISPLAY

What is an Accelerated Processing Unit?

At the most basic level, AMD’s new Accelerated GRAPHICS
Processing Units combine general-purpose x86 CPU GENERAL PARALLEL
PROGRAM DATA
cores with programmable vector processing engines on TASK PROCESSING
a single silicon die. AMD’s APUs also include a variety of ATI
STREAM
critical system elements, including memory controllers,
I/O controllers, specialized video decoders, display Designed for Designed for
outputs, and bus interfaces, but real appeal of these - General
APUs are the - Specialized
purpose tasks tasks (e.g.
chips stems from the inclusion of both scalar and vector - Most common
next generation of
graphics, video)
processors (from
hardware as full-fledged processing elements. Others applications what are known as - Most visual
have lashed a CPU and a basic graphics unit together in - Data CPUs and GPUs) applications
processing and require: - Data processing
a single package, but none have attempted this feat with serially - General-purpose, in parallel
truly programmable GPUs like those in the AMD Fusion programmable
scalar and vector
designs, let alone GPUs that can be programmed using processor cores,
high-level industry-standard tools like DirectCompute and forming a
heterogeneous
OpenCL. AMD is best situated to address this engineering compute capability
challenge, as it is currently the only company which has - High-speed bus
architecture
access to extensive IP resources (e.g. patents and - Shared, low-latency
engineering expertise) in both x86 processor technology memory model
and industry-leading GPU technology. In fact, AMD’s - All on the same die

recognition that it needed proven GPU technology for

future converged products drove its 2006 acquisition Figure 2 illustrates the arrangement of these first APUs. The
of ATI Technologies. key aspect to note is that all the major system elements –
x86 cores, vector (SIMD) engines, and a Unified Video
AMD’s APUs are set to arrive in a variety of shapes and Decoder (UVD) for HD decoding tasks – attach directly to
sizes adapted to the requirements of their target markets. the same high speed bus, and thus to the main system
AMD has disclosed that its first APUs, code-named memory. This design concept eliminates one of the
“Llano” and “Ontario,” are designed for mainstream fundamental constraints that limits the performance
desktop and notebook platforms and “thin and light” of traditional integrated graphics controllers (IGPs).
1 Subject to restrictions and limitations; fees and charges may apply.
See http://www.dell.com/us.en.business/servers/pedge_t105/pd/aspx?refid=pedge_t105&cs=04&s=bsd.
2 Additional hardware
AMD Whiteor software
Paper:mayAMD
be required
Fusionfor™remote access.
Family Internet access required.
of APUs 3
 Figure 2. APU Although the APU’s scalar x86 cores and SIMD engines
share a common path to system memory, AMD’s first
generation implementations divide that memory into
regions managed by the operating system running on
the x86 cores and other regions managed by software
running on the SIMD engines. AMD provides high speed
block transfer engines that move data between the x86
and SIMD memory partitions. Unlike transfers between
an external frame buffer and system memory, these
transfers never hit the system’s external bus. Clever
software developers can overlap the loading and
unloading of blocks in the SIMD memory with execution
involving data in other blocks. Insight 64 anticipates that
future APU architectures will evolve towards a more
Until now, transistor budget constraints typically
seamless memory management model that allows even
mandated a two chip solution for such systems, forcing
higher levels of balanced performance scaling.
system architects to use a chip-to-chip crossing between
the memory controller and either the CPU or GPU as
Just as AMD’s architects have woven x86 cores and
shown in Figure 3. These transfers affect memory latency,
GPU cores into a single hardware fabric, astute software
consume system power and thus impact battery life. The
developers can now begin to weave high performance
APU’s scalar x86 cores and SIMD engines share a common
vector algorithms into programs previously constrained
path to system memory to help avoid these constraints.
by the limited computational capabilities of conventional
scalar processors, even when arranged in multi-core
Figure 3. Typical current AMD system configuration configurations. In just a few years, machines equipped
with programmable GPUs are expected to comprise
a meaningful portion of the installed base of PCs.
Software coming from ISVs who take advantage of
these enhanced capabilities will have the ability to
execute well beyond the capability of packages that
lack support for these features.

Why all the fuss about “vector processing”?

For over four decades, architects have pursued two
competing concepts to enhance system computational
performance. One group argues that designs that
emphasize a single machine instruction operating on
a single data item (“SISD”) work with the broadest array
of application and lead to the most cost-effective
solutions. The other group counters that vector designs
Total system performance can be further enhanced that emphasize single instructions operating on multiple
through the addition of a discrete GPU. The common data items (“SIMD”) deliver relatively higher performance
architectures of the APU and GPU allow for a multi-GPU for computationally intensive applications that need to
configuration where the system can scale to harness all process large arrays of highly structured data, a model
available resources for exceptional graphics and enable that applies to many modern computing tasks.
truly breathtaking overall performance.
1 Subject to restrictions and limitations; fees and charges may apply.
See http://www.dell.com/us.en.business/servers/pedge_t105/pd/aspx?refid=pedge_t105&cs=04&s=bsd.
2 Additional hardware
4 or software may be required for remote access. Internet access required. AMD White Paper: AMD Fusion™ Family of APUs
 After all those years of debate, it has now become clear Vector processors like those used in advanced GPUs
that both sides were correct; it is not a question of “either/ have dozens, and sometimes hundreds of calculating
or,” but rather of “both/and.” Some numerically intensive units that operate simultaneously. When an application
problems lend themselves to parallel algorithms, and wants to add two one-thousand element vectors using
others don’t. When a machine optimized for parallel ten of the system’s available processing units, the vector
computation encounters a problem that cannot be software restructures the work so that each calculation
computed in a parallel manner, the machine operates executes simultaneously on ten separate elements, and
as an inefficient scalar processor, and most of its parallel thus completes the work in as little as one-tenth the time.
computing resources sit idle. Conversely, a processor Sounds easy, doesn’t it? Of course, one must account
optimized for scalar calculations cannot exploit the for the time needed to set the operations up, the time
parallelism in many algorithms, and thus is limited by needed to ensure they all complete successfully, and the
its scalar processing speed. time needed to move data between the system’s scalar
and vector memory regions. One must also ensure that
Affordable PCs that offer hundreds of gigaFLOPS the operations applied to any element in each vector can
performance are not equipped to cause spreadsheets to be performed independently of operations applied to
recalculate faster or e-mail to get to its destination sooner, other elements in that same vector. It is easy to see that
but they can impact the way users interact with their implementing these features in hardware might be more
system to set up those spreadsheets or compose that complicated than one might initially think.
mail. A new era of Natural User Interfaces (NUIs) can
help users to communicate with their system using visual Given the restrictions and overheads enumerated above,
and verbal inputs instead of mice and keyboards. Step- along with a few others, it should come as no surprise
function increases in system capabilities and usability that vector processing techniques cannot boost the
have a way of shaking up the status quo in existing performance of all programs equally. For small data
markets, benefiting those who recognize and react to arrays, the overhead associated with setting up vector
the impending changes, and negatively impacting those operations can outweigh the time saved through parallel
who fail to recognize the shift, or react too slowly to it. execution. Many problems and algorithms have proven
a poor fit for this technology, and are best handled
The ABCs of Vector Processing using scalar approaches. Some early vector-oriented
The technology that allows vector processors to achieve architectures excelled on vector workloads but failed
very high levels of computational performance is easy in the market because their designers neglected these
to understand, but difficult to implement. This partially scalar workloads.
explains why so few companies have fielded successful
products based on this approach. Scalar processors AMD’s Accelerated Processing Unit designs are
operate on arrays of data one element at a time. When a constructed to help avoid these pitfalls. These APUs
program needs to add one thousand elements in vector A cut through scalar workloads using AMD’s proven x86
to a separate list of one thousand elements in vector B, core technology and through vector workloads using
and store the results in vector C, it typically sets up enhanced versions of its GPU technology. Although
pointers to each vector, loads the values pointed to by A AMD had to overcome many technical challenges to
and B into CPU registers, adds those registers, and stores merge its vector and scalar technologies in a manner that
the result into the location pointed to by C. Then the preserves the advantages of each, having the core IP for
program updates all its pointers and repeats the process both processing elements provides this AMD hardware
one thousand times. The actual time the CPU spends with a significant advantage over other hardware designs
handling the one thousand “add” operations pales in which are missing one or the other.
comparison to the time it spends handling the looping
operations – updating pointers and iteration counts.
1 Subject to restrictions and limitations; fees and charges may apply.
See http://www.dell.com/us.en.business/servers/pedge_t105/pd/aspx?refid=pedge_t105&cs=04&s=bsd.
2 Additional hardware
AMD Whiteor software
Paper:mayAMD
be required
Fusionfor™remote access.
Family Internet access required.
of APUs 5
 How hard is it to accelerate software using ISVs who focus on Windows and use DirectX APIs for
GPU computing resources? graphics will most likely gravitate toward DirectCompute.
The recent emergence of two important development It features data structures compatible with DirectX 10
tools – OpenCL and DirectCompute – enables developers and DirectX 11 application programming interfaces (APIs),
(especially those who have already mastered the art of and thus helps simplify the process of adding GPU
writing software for single-threaded scalar environments) acceleration for physics or Artificial Intelligence tasks to
to more easily create highly dynamic multi-threaded DirectX 11 applications. Microsoft has further encouraged
data-parallel software applications. this by releasing DirectX 11 and DirectCompute APIs as
“platform extensions” to Windows Vista, which allows
Over the past three years, the tools AMD provides to software that uses these APIs to target a huge installed
developers who want to use GPU hardware to accelerate base of systems, along with all the new Windows 7
their applications have undergone a dramatic evolution. systems being sold now.
Prior to November 2006, ISVs were limited to applying
ATI Radeon™ graphics cards or ATI FirePro™ graphics ISVs wishing to address a broader market may find
accelerators to anything other than traditional 3D OpenCL the better development platform, especially
applications. Then AMD launched the “Close To Metal” if they already use OpenGL APIs to handle graphics
(CTM) initiative that gave early GPU-computing pioneers interfaces in existing applications. The Khronos Group,
a set of low-level, proprietary interfaces they could use to an industry consortium that drives a variety of open
craft GPU-accelerated applications like Folding@Home, a API standards, serves as the development coordinator
distributed computing project run by Stanford University for both OpenCL and OpenGL languages. OpenCL
that puts the idle time normally wasted by screen savers includes support for both data-parallel (i.e., SIMD) and
in millions of PCs to good use in the study of protein task-parallel execution models. It uses data structures
folding, aggregation, and related diseases. compatible with OpenGL APIs, thus helping to simplify
the task of adding GPU compute acceleration to OpenGL
Roughly a year later, AMD augmented its GPU software applications. Key processor suppliers, including AMD,
toolkit with the release of the ATI Stream SDK and Brook+, ARM, Intel and IBM, and key GPU suppliers, including
an open-source, high-level C-like language that simplified AMD and Nvidia, support OpenCL and have released
the use of AMD GPUs for computational tasks. This OpenCL drivers that work in a number of OS environments
brought a few more developers on board the GPU- for many of their devices. AMD’s OpenCL compiler
computing bandwagon, but most ISVs held out for supports both its ATI Radeon™ HD 4000 and HD 5000
industry-standard based tools that would allow them series GPUs and its multi-core x86 processor offerings.
to address a broader range of hardware configurations.
That wait officially ended last fall, with the release of ISVs who want to get a head start on developing
OpenCL, a cross-platform standard for parallel computing software for AMD’s forthcoming APU offerings can begin
coordinated by the Khronos Group; and DirectCompute, their work today on any AMD platform that includes an
a new set of Windows DirectX APIs that facilitates GPU ATI Stream enabled ATI Radeon HD 4000 or HD 5000
computing applications. Thus ISVs have moved in short series discrete GPUs. The new APUs will appear just
order from having no industry-standard options for GPU like today’s DirectCompute and OpenCL platforms from
computing to having two. the viewpoint of the software that runs on the platform.
This means investments ISVs make today to support
current platforms will remain useful on future platforms
addressing a much wider breadth of the PC marketplace.

1 Subject to restrictions and limitations; fees and charges may apply.

See http://www.dell.com/us.en.business/servers/pedge_t105/pd/aspx?refid=pedge_t105&cs=04&s=bsd.
2 Additional hardware
6 or software may be required for remote access. Internet access required. AMD White Paper: AMD Fusion™ Family of APUs
 Does this stuff really work? »»Another startup has demonstrated facial recognition
The tools needed to accelerate applications via GPU software that finds individual faces in photos or videos
computing have only been around for a few years, and matches them to faces in its database. This obviously
but already a few innovative ISVs have used this requires a tremendous amount of computational
technology to enhance their applications. A few of horsepower, but with GPU assistance, it can accomplish
the more interesting examples include: this task virtually in real time. It’s not hard to imagine that
GPU computing could be employed by civil protection
»»Adobe’s ubiquitous Flash Player now uses GPU
organizations to help make the world a safer place.
hardware to decode video streams. This innovation
helps improve the quality of the video playback on
What can APUs do for traditional workloads?
enabled GPUs, reducing the processing load on the
Although it’s exciting to look at the new applications that
CPU, and thus uses less power, extending system
will finally become practical in the “Fusion” era, the fact
battery life. The release candidate of Flash Player 10.1
remains that most users will want their new APU-based
is available for download as of this publication date.
systems to handle a mix of traditional applications for
»»ArcSoft has added a GPU-enabled SimHD™ plug-in office productivity and Internet access, along with those
to its TotalMedia Theatre package. The new plug-in new exciting apps. Fortunately, the changes AMD made
enhances video quality by intelligently up-scaling to enable new APU-accelerated applications can also
standard DVD video from 480 vertical lines to 720. help existing applications run better as well.
»»Cyberlink has enhanced its line of media software
applications to use GPU acceleration whenever Many of these improvements stem from AMD’s ability to
possible. Its Power Director 8 package takes advantage fit the CPU cores, GPU cores and North Bridge (the part
of enabled GPU resources to speed up video editing, of the chip where the memory controller and PCI-express
video encoding and video effects rendering. Its interfaces reside) onto a single piece of silicon. As noted
MediaShow line uses enabled GPU hardware to earlier, this eliminates a chip-to-chip linkage that adds
accelerate video format conversion (transcoding) and latency to memory operations and consumes power. It
encoding, as well as to implement an automated “face takes less energy to move electrons across a chip than
tagging” feature that sorts the user’s photo collection to move those same electrons between two chips, and
based on the faces in the photos. Its PowerDVD offering the power saved by this small change alone can help
takes advantage of GPU resources to enhance Blu-Ray significantly increase system battery life. The co-location
playback; the company demonstrated a future version of all key elements on one chip also allows AMD to take a
of PowerDVD for Blu-Ray 3D playback at the 2010 CES holistic approach to power management on these APUs.
show. Cyberlink started its GPU-acceleration efforts in They can power various parts of the chip up and down
2008, and used the proprietary tools then available depending on workloads, squeezing out a few milliwatts
from AMD and Nvidia to develop its software. Now it is here and another few milliwatts there – which in the
converting its software to use DirectCompute in order to aggregate can amount to significant power savings.
increase the range of supported platforms and get to
market with new features more quickly. Finally, some of the improvements can be attributed to
the advanced GPU technology AMD embeds in its APU
»»One Silicon Valley startup uses GPU resources to clean offerings. Although the company has yet to reveal the
up video files, compensating for noise, pixilation, graininess, technical specs of these GPUs, it has disclosed they will
poor focus, low contrast, and shaky images due to shaking be DirectX 11-compliant. These will be the first APU-based
cameras. The package works just like the fictional ones you systems that can support DirectX 11’s enhanced visual
might see in a film where the hero zooms in on a satellite experience without a discrete GPU, and thus will represent
image and reads the villain’s license plate, but this package a cost-effective solution for systems developers.
relies on GPU hardware, rather than Hollywood gimmicks.

1 Subject to restrictions and limitations; fees and charges may apply.

See http://www.dell.com/us.en.business/servers/pedge_t105/pd/aspx?refid=pedge_t105&cs=04&s=bsd.
2 Additional hardware
AMD Whiteor software
Paper:mayAMD
be required
Fusionfor™remote access.
Family Internet access required.
of APUs 7
 Summary: What can APUs do for PC users? Since the days of the earliest personal computers, each
The AMD Fusion family of Accelerated Processing Units major advance in system capability has enabled
is scheduled to arrive in 2011. The expectation is that their innovative software developers to create new products
compatibility with Windows 7 and DirectX 11 will ensure that opened new markets. The Apple II gave us VisiCalc,
that they will provide an outstanding experience for those the first spreadsheet. The original IBM PC led to Lotus 1-2-3,
who purchase PCs based on these processors. Their the first spreadsheet with graphics. The Macintosh ushered
enhanced processing power and power efficiency will in an era of desktop publishing that has forever changed
enable sharp and clear videos, realistic and responsive the way the world creates and distributes information.
games, and notebooks that can run longer between
battery charges. The dramatic increase in performance enabled by
AMD Fusion technology can create new opportunities
More importantly, compared to today’s mainstream for entrepreneurial developers to innovate and make the
offerings, APU-based platforms will possess prodigious world a better and richer place. Along the way, they may
amounts of computational horsepower. This processing enrich themselves as well. That’s the way the system is
power will allow developers to tackle problems that lie supposed to work.
beyond the capabilities of today’s mainstream systems,
and will enable innovative developers to step up and For the hardware developer, ODM or PC manufacturer,
update existing applications or invent new ones that it’s time to start thinking about how to incorporate these
take advantage of GPU acceleration. These features new APUs into product lines in order to enhance the
will be a standard part of every APU. Over time, even consumer experience. Software developers should look
the most affordable PCs can be expected to have the to this new power to help their software run even better.
computational performance of yesterday’s million dollar All developers are encouraged to upgrade their skills and
mainframes with “all day” battery life. learn about OpenCL and DirectCompute, and to examine
current software projects to see how they can be
Of course, few users will want to run the same applications improved in a world where systems have dramatically
on tomorrow’s notebooks that they ran on yesterday’s more power. Because pretty soon, they will.
mainframes and supercomputers. They will likely want
to run applications that help them in their everyday lives,
doing tasks they cannot accomplish on the systems
they own today. They may want to use facial recognition
software to sort their photos and videos, or even to help
them identify people they meet on the street or actors
they see in movies. They may want the on-screen
appearance of the videos they stream to approach that
of the HD content on their TVs, even when bandwidth
constrains that content to a low resolution format.

©2010 Advanced Micro Devices, Inc. All rights reserved. AMD, the AMD Arrow logo, ATI, the ATI logo, FireGL,
FirePro, Radeon and combinations thereof, are trademarks of Advanced Micro Devices, Inc. Microsoft,
DirectX and Windows are registered trademarks of Microsoft Corporation in the United States and/or other
jurisdictions. OpenCL is a trademark of Apple Inc. used by permission by Khronos. Other names are for
informational purposes only and may be trademarks of their respective owners.

48423B