Page 1
GE
Intelligent Platforms
GPGPU
COTS Platforms
High-Performance Computing Solutions
Page 2
Rugged GPGPU COTS Boards
for Military and Aerospace
GPGPU platforms deliver new levels of
performance for size, weight and power
(SWaP) constrained mission payloads.
PARALLEL COMPUTING
2 cores
Optimized for throughput computing
The world of high-performance computing
is undergoing a revolution, thanks to
advances in General Purpose computing
on Graphics Processing Units (GPGPU). The
idea behind GPGPU is to use a GPU, which
typically handles computation for computer
graphics only, to perform parallel computation in applications that have traditionally
been handled by the CPU.
A multi-GPU platform hosted by one or
more CPUs is able to perform heterogeneous computing, harnessing the parallel
computing power of the many-core GPUs to
provide very large increases in performance
with minimal programming complexity.
96 cores
Additionally, programmers are helped with
software development environments such
as Compute Unified Device Architecture
(CUDA) and OpenCL, which allow them to
harness the many-core, parallel processing
capabilities of the GPGPU platforms.
While greatly increasing functional capability, the GPGPU platform also delivers the
performance with far less size, weight and
power (SWaP). This results in significant
savings in cost, risk, and time-to-market.
Lab-proven technologies ruggedized for the rough and tumble
world of military applications
Now these benefits are fully available for
rugged military and aerospace applications.
With the introduction of a full range of GE
Intelligent Platforms rugged GPGPU boards
and systems, the advantages of GPGPU are
no longer confined to controlled environments
at universities, research centers and hospitals.
The unique partnership between GE Intelligent Platforms and NVIDIA allows for new
product development using NVIDIA GPUs
based on the award-winning CUDA architecture, for Military and Aerospace applications.
2
gpgpu • www.ge-ip.com/gp gpu
Page 3
Enhanced Performance for
1
1
Far Less Size, Weight and Power
The graph shows GFLOPs performance
for the VSIPL multiple FFT operation. It
compares performance on different GE
Intelligent Platforms embedded platforms;
a GPU platform, multi-core Intel Penryn and
i5 platforms and an e600 based PowerPC
platform. GE’s platform-optimized AXISLib
DSP library product is used on all platforms.
Data is in GPU memory for the GPU case
and in-cache for the Intel and PPC cases.
MULTIPLE FFT PERFORMANCE - GPU vs INTEL
60.00
50.00
40.00
GFLOPs
30.00
20.00
10.00
0.00
000256
000256
001024
000100
004096
000050
MATRIX SIZE
016384
000020
065536
000020
INCREASING PERFORMANCE ADVANTAGE OF GPU
PEAK MEMORY BANDWITH (GB/sec)PEAK PERFORMANCE (Gflop/Watt)
0
GPU
CPU
8
6
4
2
0
2003 2004 2005 2006 2007 2008 2009 2010
Magic1 – G73
G70
Pentium M
GeForce 240M
Core 2 Quad
GeForce 480M
G92
Core i7
00
80
60
40
20
0
2003 2004 2005 2006 2007 2008 2009 2010
GPU
CPU
Harpertown
3 GHz
131072
000020
GT240/CUDA VSIPL
Penryn - 2 threads
i5 - 2 threads
i5 - 4 threads
PPC - 1 thread
GeForce 8800M
GeForce Go 7800
Nehalem
3 GHz
GE’s new GPGPU platforms are particularly well-suited to many of the processing
SIZE WEIGHT POWER
needs of Military and Aerospace applications where size, weight and power are key
considerations along with resistance to
extended temperature, shock and vibration.
GPGPU technology allows system designers
to pack more punch into less space and use
less power for your applications.
Industrial Standard
(576 Gflops)
GPGPU Products
(766 Gflops)
4 cu 105 lbs 2000 W
0.8 cu
10 lbs
200 W
REDUCTION IN:
• COST
• RISK
• TIME TO MARKET
3
Page 4
Rugged GPGPU COTS Boards
for Military and Aerospace
Radar
One of the biggest challenges for today’s
radar systems is to provide more capability—range, number of targets, speed,
etc. —while meeting ever more stringent
SWaP constraints. The extra speed offered
by the GPGPU platforms translates directly
to more area coverage and more security
for the operating team.
One rack containing 72 conventional
processors (18 6U boards) and producing a
peak capability of 576 GFLOPS can take up
4 cubic feet, weigh over 105 pounds and
consume over 2000 watts. GPGPU tech-
nology can allow system designers
to fit an unprecedented amount
of processing power into a
very compact package.
The use of three 3U VPX
boards can yield peak
processing power of
766 GFLOPS in less
than 0.4 cubic feet.
Development Ease
Increases in performance will be obtained in
application areas such as Software Defined
Radio, sonar, and medical imaging. But what
is less obvious is the change in development
strategy offered by GPGPU technology. The
only other technology currently offering
massively parallel processing capability is
Field Programmable Gate Arrays (FPGAs).
Although FPGAs provide very highperformance data processing, developing
high-performance FPGA cores requires a
very specialized skill set built on a hardware
engineering background, whereas developing code for GPGPU processors is much
more of a software issue. For companies
with a background in multi-processor GPP/
DSP-based system architecture, the move
to GPGPU will be much less disruptive than
a move to FPGA processors. The processing
power, system size and power consumption
are compelling factors, but the addition of
programming ease makes such a system
tough to match.
4
Page 5
Data Encryption/
Decryption
There are several standards for encryption
of data, including Advanced Encryption
Standard (AES). AES is the first publicly
accessible and open cipher approved by
the U.S. National Security Agency for topsecret information, typically requiring 256
bit keys at this level. The time to encrypt a
block of data increases linearly with the size
of the key. The computation load required
to maintain encryption of a real-time data
stream can be prohibitive. With the advent
of CUDA and the addition of crucial arithmetic, bitwise logical and shift operations as
well as the ability to use texture caches to
index tables, GPUs are now a viable option
to general-purpose processors for data
encryption/decryption. Performance gains
up to 10x have been demonstrated.
Situational Awareness
Surveillance of large areas has historically
been achieved by using an array of sensors
connected to a bank of monitors, with separate or multiplexed displays for each video
stream. Such arrangements present the
operator with a confusing array of disparate
video feeds, require a great deal of space,
and consume a large amount of power. In
a dynamic, real-time scenario, there is also
a danger of information overload for an
operator attempting to interpret such large
volumes of imagery. Interrelationships between sensors is not
always obvious, and
important contextual
visual information
can be overlooked.
Many such
systems rely on
the operator for
“event” detection, but large volumes of information coupled to the effects of stress and
fatigue can significantly reduce operator
effectiveness.
Our image processing subsystem overcomes these issues and greatly improves
the performance of surveillance assets
and their operators. We offer a previously
unattainable level of situational awareness
to platforms such as armored vehicles,
aircraft, remote unmanned platforms and
security and surveillance systems.
5
Page 6
IED Detection
Target Tracking
Improvised Explosive Devices (IEDs) are a
major cause of injuries and fatalities among
ground troops. A number of techniques for
automated detection of IEDs are used, and
all of these require processing a high volume
of data. The effectiveness of the solution
depends on how fast the algorithm can reliably operate on that data. GPGPU technology
is proving to be a highly effective solution for
such high-throughput computations.
Ground Change Detection relies on realtime image processing, and may be applied
to sensors mounted on ground vehicles or
UAVs. The system needs to apply image
registration and stabilization, and moving
object extraction, before comparison with
normalized geo-referenced data, all while
dealing with lighting and legitimate scene
changes. Ground Penetrating Radar (GPR)
allows construction of a 3D model of the
ground, identifying any suspicious objects
or changes from normalized data. GPR can
be applied to ground mobile or airborne
systems. In addition, behavioral modeling—
based on live sensor imagery possibly
combined with wide-area surveillance
data—can be used to identify hostile intent
and potential threats, giving operational
forces time to assess the risk and take
appropriate defensive actions.
GPGPU-based automatic video trackers
and image processors are at the heart
of target tracking systems where they
provide the highest performance solutions
in the smallest, fully ruggedized hardware
packages. Target detection and target
acquisition processes identify objects within
an area of the video image display that
meets the user-defined target criteria. A
range of detection algorithms are built into
the system to meet situational requirements. When one or more targets have
been detected, the tracking system can
enter automatic or manual tracking mode.
Automatic target acquisition may be prioritized by using several different factors, such
as target nearest to the boresight or the
largest target. If a system is in autotrack
mode, the video tracker automatically
tracks the selected target and can control
almost any type of pan-and-tilt or gimbal
system to track the target. GPGPU-based
automatic video trackers feature a wide
range of proven high-performance detection and tracking algorithms that can
be tailored to the operational scenario,
including centroid, phase correlation
and edge detection. Algorithms may be
combined to meet particularly demanding
tracking scenarios.
6
gpgpu • www.ge-ip.com/gp gpu
Page 7
GPGPU COTS Platforms
GE Intelligent Platforms GPGPU-empowered
platforms can be easily implemented to
either adapt to your legacy applications or to
accommodate your new applications.
LEGACY
SOFTWARE
YES
MAXIMUM
PERFORMANCE
DESIRED?
NO
ADD CUSTOM
CUDA KERNELS
(GE consulting)
APPLICATION CODED
WITH
AXISLIB-GPU VSIPL
AXISView
PERFORMANCE
TUNING
AXISFlow
I/O INTEGRATION &
SYSTEM INTEGRATION
DEVELOPED
SOLUTION
7
Page 8
Application Development
Made Simple By Software
AXISLib-GPU provides the industry standard
Vector, Signal and Image Processing Library
(VSIPL) for NVIDIA-based GPU platforms. The
VSIPL API model fits perfectly with GPGPUenabling DSP application developers to
quickly realize significant application
speed-ups without the need to learn GPU
programming techniques.
AXISLib-GPU supports the development
and speeds the deployment of highperformance DSP and multiprocessing
applications on GE’s NVIDIA CUDA-enabled
GPGPU platforms such as the IPN250,
NPN240 and GRA111. Typical applications
include radar, sonar, image processing,
signals intelligence and intelligence, surveillance, reconnaissance (ISR). It is a set of
signal and vector processing libraries
providing over 500 high-performance digital
signal processing and vector mathematical
functions optimized for NVIDIA’s many-core
GPUs and created to help developers maximize system and application performance.
INTEGRATED SOFTWARE MODULES
Multiprocessing Application
AXIS Advanced Multiprocessor Integrated Software
Optimized
Math & Function Libraries
VSIPL, GE 542+ functions
PowerPC/Altivec, Intel SSE,
Generic C, CUDA
Board Support Package and Drivers (VxWorks, LynxOS BSP)
PPC/Intel/GPU PCI/Starfabric/PCIE/SRIO/10GE
Inter-Processor
Communication
Starfabric, PCI, PCIe,
SRIO, TCP/IP
Universal Interface Layer (UIL)
Productivity Suite
(performance tuning)
AXIS
Middleware
Abstraction
Layer
Traditional
Hardware
Support
The AXIS Multiprocessing software suite
facilitates the development of complex
applications over multiple clusters of GPU
platforms.
AXISFlow provides a communications API
for multi-threaded/multi-core/multiprocessor communications.
AXISView provides a set of GUI tools
enabling system visualization, application
instrumentation, debug and monitoring.
8
gpgpu • www.ge-ip.com/gp gpu
A view of 2 x IPN250s
and an NPN240 system
using the AXISView
graphics tool
Page 9
PROCESSING FLOW ON CUDA
CUDA exploits the massively parallel characteristics of NVIDIA’s ubiquitous silicon.
CUDA is taught in universities worldwide
and used in many R&D labs, so a large
number of programmers are available and
there is a wealth of web-based resources.
CUDA software development tools:
• NVIDIACCompilerforparallelGPUcode
• CUDADebugger
• CUDAVisualProfiler
• SDKwithbest-practiceguides
• ParallelNsight
®
Copy processing
data
Main
Memory
Memory
for GPU
Copy the result
GPU (GeForce 8800)
CPU
Instruct the
processing
Execute parallel
in each core
Advanced libraries that include:
• NVIDIAPerformancePrimitives(image
and video)
• BasicLinearAlgebraSubprograms
• FFT
• VSIPL
GE GPGPU products also support Open
Computer Language (OpenCL), the first
open language for writing programs that
execute across CPUs, GPUs, and other
processors. It includes a language for
writing kernels, defines APIs, and provides
parallel computing using task-based and
data-based parallelism.
Open Graphics Library (OpenGL) is a
standard specification defining a crosslanguage, cross-platform API for writing
applications that produce 2D and 3D
computer graphics. This is used in the
graphics output processes.
C for CUDA extends C by allowing the programmer to define C functions, called kernels, that
when called are executed N times in parallel by N different CUDA threads, as opposed to
only once like regular C functions.
A kernel is defined using the __global__ declaration specifier and the number of CUDA
threads for each call is specified using a new <<<…>>> syntax:
__global__ void vecAdd(float* A, float* B, float* C)
{
int i = threadIdx.x;
C[i] = A[i] + B[i];
}
int main()
{
//Kernelinvocation
vecAdd<<<1, N>>>(A, B, C);
}
9
Page 10
COTS GPGPU
Boards and Systems
GRA111
NVIDIA GT 240 CUDA Capable Graphics Processer
The OpenVPX-compatible GRA111 is the second generation 3U VPX
graphics board, employing the state-of-the-art NVIDIA GT 240 GPU
to bring desktop performance to the rugged Military and Aerospace
market. In addition to meeting increased demand for graphics
rendering performance, the GRA111 is the first rugged implementation of a CUDA capable GPU.
The GRA111 supports the 16-lane PCI Express implementation,
providing the maximum available communication bandwidth to a
CPU such as GE Intelligent Platforms SBC341. The PCI Express link will
automatically adapt to the active number of lanes available, and so
will work with single board computers in 4- and 8-lane configurations.
SBC341
VPXcel3, Penryn
The SBC341 is a 3U VPX Single Board Computer based on the Intel
Core 2 Duo (Penryn) processor and is part of the VPXcel3 product
family. This processor is 45nm micro architecture delivering superior performance and energy-efficiency.
This single board is optimized for use in VPX systems with multiple
PCI Express port options including an x16, x4, and x1. The x16 is
designed to provide a dedicated communication path to a graphics
processing unit, such as the GRA110. The x4 and x1 PCI Express
ports allows high-speed communication to other single board
computers and I/O cards in order to build complex systems.
10
gpgpu • www.ge-ip.com/gp gpu
Page 11
IPN250 and NPN240
Intel + NVIDIA
These OpenVPX CPU+GPU combination boards feature NVIDIA CUDAcapable GT240 96-core GPUs, where the IPN250 uses 1 NVIDIA GPU
and the NPN240 uses 2 NVIDIA GPUS. The GPUs enable the boards
to be 15 to 20 times more energy efficient than CPU boards. Multiple
boards can be linked to single or multiple hosts to create multi-node
CUDA GPU clusters capable of thousands of GFLOPS.
Main Features
• PCIeGen2fabricinterfaces
• Multiplecaptureandoutputformatssupported
• OpenVPX/REDI-6Ruggedizationbuildlevels
SBC324 and SBC624
2nd Generation Intel Core i7
The VPXcel3 SBC324 is a 3U VPX Single Board Computer built around
the 2nd Generation Intel Core i7 processor with fully integrated
graphics and memory controller in one device. This 32 nm processor
incorporates Advanced Vector Extensions (AVX) technology that
provides tremendous signal processing capability in a 3U form factor.
Coupled with the QM67 chipset, the rugged OpenVPX SBC324 is
available with dual- or quad-core processing at up to 2.5 GHz.
Getting Started
TheCUDAStarterKitisintendedtoallow
engineers a quick and easy start up with a
CUDA-based system, either for evaluation
and benchmarking or to develop a specific
application. Containing all the hardware
that is needed to begin development,
customers can immediately begin loading
their chosen operating system on the delivered equipment.
NVIDIAdriversandCUDASDKaredeliveredonCDfor
Wind River Real Time Linux
time Linux, Red Hat, SuSE, Windows
Vista. For latest versions and technical support, please
refer to www.nvidia.com
®
, Concurrent Red Hawk real
®
XP, and Windows
The VPXcel6 SBC624 Single Board Computer is also based on the
2nd Generation Intel Core i7 processor and QM67 chipset. The
SBC624, which offers a range of air- and conduction-cooled build
levels, provides an unmatched level of I/O bandwidth for both
onboard and offboard functions. Onboard XMC mezzanine expansion sites provide enhanced system flexibility.
SC-S-CUDA3USK1
3U GPGPU Development Chassis
MAGIC1
Pre-configured Rugged 3U GPGPU Development Chassis Solution Set Contains:
• SBC341IntelCPU
• GRA111NVIDIAGPU
• 64GBsolidstateSATAdrive
SC-S-CUDA6USK1
6U OpenVPX Development Chassis
Supports multiple fabric topologies within a single
chassis without having to define the system backplane
requirements in advance of application development,
which greatly reduces risk and time to solution for your
applications
11
Page 12
GE Intelligent Platforms Contact Information
Americas: 1 800 433 2682 or 1 434 978 5100
Global regional phone numbers are listed by location on our web site at www.ge-ip.com/contact
defense.ge-ip.com/gpgpu
©2011 GE Intelligent Platforms, Inc. All rights reserved.
*Trademark of GE Intelligent Platforms, Inc.
All other brands or names are property of their respective holders.
06.11 10M GFA-1790
Loading...