You are on page 1of 11


White Paper

Meeting Military Data Signal

Analysis Imperatives
Radar, Sonar and Image Processing
In 21st century electronic warfare, one of the most decisive
factors behind victory is the ability to aggregate massive
amounts of data and process those streams into insightful,
actionable outcomes. Data processing can distinguish
between noise and a viable target, and the speed of that
processing can mean the difference between real-time
intelligence and missed opportunities.

Table of Contents
Obstacles in Military Signal Processing P2 SWaP Considerations in Military
The Demand for Increased Processing Power P2
Signal Processing
Picking the Processor - why GPGPUs P3
ADLINK Product Offerings With unlimited size, weight, and power consumption (SWaP),
signal processing solutions could achieve almost unlimited
Radar Processing P5
results. Of course, SWaP limitations are a reality in practical
Sonar Processing P7 applications, particularly in increasingly mobile military
Image Processing P8 environments. SWaP discussions necessarily revolve around
Conclusion P9 platform efficiency: How much performance can be obtained
About ADLINK P9 from a platform on a given power and/or cost budget?

Obstacles in Military Signal Processing - Speed Bump
Over 15 years ago, Intel and other CPU manufacturers realized that the race for raw speed, as measured in megahertz,
was destined for failure. At roughly 4 GHz, processors encountered a ceiling at which heat began to cripple performance.
Multi-core processors arose to address the problem, essentially harnessing multiple smaller engines that could collectively
outperform a single large engine. GPU chips soon followed CPUs in the shift to efficiency over speed, with NVIDIA starting
around 2010 to discuss performance in terms of gigaFLOPS (floating point operations per second) per watt.

The 2016 BERTEN white paper “GPU vs FPGA Performance Comparison” provides an intriguing example of such efficiency
breakdowns. The authors ran single precision floating-point performance benchmarks on six processor options. Two of the
three GPU options handily beat every FPGA contender. However, when examined on a gigaFLOPS (floating point operations
per second) per watt basis, FPGAs were clearly the preferable option. And yet, when reassessed for price efficiency (cost per
gigaFLOPS), GPU numbers were a small fraction of their FPGA counterparts. The clear take-away was that while in 2016,
FPGAs were the most power efficient, GPUs delivered by far the best performance for the money. This should be relevant for
any organization needing to advance its signal processing capabilities under a constrained budget.

Not surprisingly, component and solution providers continue to hunt for every path possible to reduce power demands.
While NVIDIA started to discuss performance in terms of gigaFLOPS per watt in 2010, the company has more recently taken
to analyzing these metrics within specific applications, workload types, and system architectures. For example, NVIDIA
provides twin sets of specifications for its Tesla V100 GPU, depending on whether the host system uses NVLink or PCI
Express bus connectivity. With PCIe, the V100 has a maximum power draw of 250W, but that power draw is sufficient for 7
TeraFLOPS performance in double-precisionbenchmarks or 112 TeraFLOPS in deep learning. In contrast, NVLINK has a top
power envelope of 300W, but it offers higher bandwidth and lower latency and thus the ability to deliver higher efficiency.
Today, the Intel server architectures prevalent throughout signal processing applications only allow for PCI Express links
between CPU and GPU, so this model dominates modern solution architectures. Regardless of the specifics, though, expect
power/performance efficiency to remain a key factor in GPGPU platform characteristics and selection.

The Demand for Increased Processing Power

Power also dictates some aspects of solution size. Throughout the military, missions increasingly revolve around command,
control, communications, computers, intelligence, surveillance, and reconnaissance (C4ISR) capabilities being stuffed into
vehicles ranging from bombers to scouts. Demands for nimbleness and mobility dominate field work, and the weighty
rackmounted, generator-driven systems of yesterday are quickly giving way to small form factor (SFF), battery-optional
solutions able to fit under a seat. Embedded platforms are becoming the norm, but they must be able to meet the performance
demands of modern signal analysis work. Concurrently, those smaller-footprint enclosures face greater challenges with
dissipating component heat resulting from more rigorous computing workloads. With no room for big fans, and the need for
designs that can withstand the dust, shocks and vibrations of field use, passive cooling through conduction becomes

Traditional approaches to military signal processing have struggled with the dilemma of either performing analysis on-site
and dealing with the limitations of field systems, or sending data back to more capable, centralized systems located far from
the field of operation, which then incur communication latencies. The strain on both approaches only magnifies as modern
sensor systems evolve in complexity and capacity. Consider electro-optical and infrared (EO/IR) solutions, such as WESCAM’s
MX-25D imaging sensor for airborne targeting. The MX-25D incorporates up to nine sensors spanning thermal, optical zoom,
HD low-light spotting, and laser tracking capabilities. The amount of data output from this single unit is formidable, but bear
in mind that these are 720p and 1080p cameras. Imagine the data load from next-generation sensors supporting 4K imaging
or 360-degree virtual reality capture systems that might incorporate well over a dozen cameras per unit. Depending on the
application and situation, several such sensor clusters might be pooled into a central data pool for analysis, compounding the
analysis workload many times over.

Picking the Processor - Why GPGPUs?
Processors must cope with this data deluge, but the type and number of processors in play can make every difference in
analysis performance and the solution’s ultimate viability. Generally speaking, processors for such tasks fall into three groups:

Central processing units (CPUs)

also known as general-purpose processors (GPPs). In the server arena, The Intel® Xeon® processor family has long dominated
the field. These CPUs excel at tackling a broad range of task types in an ad hoc or randomized fashion.

General purpose graphics processing units (GPGPUs)

arose from 1970s arcade gaming roots. In the early 2000s, NVIDIA and then ATI developed methods for running small on-GPU
programs for individual pixel computations. This quickly evolved into parallel “stream processing” across many logical cores
within the GPU. NVIDIA launched its CUDA programming model in 2007 to help developers gain easier and more effective
access to general purpose GPU (GPGPU) computing, which provided several times faster performance on many highly
parallelized functions compared to what CPUs could manage.

Field-programmable gate array (FPGA)

processors are designed to be configured by customers or integrators subsequent to manufacture (thus being
“field-programmable”). The many logic blocks within FPGAs can perform complex combinational problems, allowing them to
effectively tackle virtually any problem. As such, FPGAs can be extremely efficient for specific computing tasks, including
parallelized operations, but they remain relatively difficult to program and involve longer development schedules than
CPU- or GPGPU-based applications.

All three of these processor types are used extensively in military applications, but for real-time signal processing and
analysis, GPGPU solutions offer the best combination of value and benefits in increasingly diverse scenarios. Other GPGPU
advantages include:

High video memory bandwidth to keep large problem computation closer to the main processor
Comparatively strong floating point computation performance
A multi-core architecture in which most chip silicon is devoted to computing units, not cache,
which further aids highly parallelized data processing
Relatively easy programming via high-level languages (CUDA, C, C++, Python, etc.)
OpenACC and OpenCL support for shorter iteration cycles in algorithm transplants
Higher computing power efficiency relative to CPU-only computing

In short, GPGPU architectures are ideal for applications that benefit from massively parallel processing of large data volumes
and/or arithmetic-intensive computing, particularly involving multiple calculations based upon a single memory visit. Highly
arithmetic-intensive workloads handled by many data processing units mean that memory access latencies can be
compensated for through accelerated computation without the need for extensive data caching. According to NVIDIA, these
architectural advantages can yield 10X to 100X advantages over CPU-only computing in applications such as computer vision,
password deciphering, and controlled simulation. Such benefits come with challenges, of course, and these must be weighed
when considering military signal analysis solutions.

Military solutions for signal analysis must address all of these factors within the context of given applications. Taking GPGPUs
as the most efficient solution available today for signal analysis, both in terms of raw performance and power dynamics, we will
now examine three particularly promising areas for military GPGPU adoption: radar, sonar, and image processing.

ADLINK Product Offerings
Radar Processing
Synthetic aperture radar (SAR), phased array radar, and hybrid
radar systems are prevalent in military information gathering.
Applications include air-defense systems, antimissile systems,
aircraft anti-collision systems, ocean surveillance systems,
altimetry and flight control systems, and guided missile target
locating systems. The suitability for GPGPU acceleration in this
space dates back nearly a decade. Consider the benchmarking
results of Peter Morris, et. al.,from India’s Defense Research &
Development Organization, which compared an Intel® Xeon®
processor-based system with eight NVIDIA Quadro FX 3800
GPUs (launched in 2009) against a PowerPC equivalent. The
GPGPU-driven system realized acceleration rates for various
radar tasks (data conditioning, MTI, Doppler processing, etc.)
from 16X to 82X.

Consider just one example of current developments in military

radar applications: The U.S. Navy recently announced a $3
million contract to design a GPGPU-based upgrade to the
Lockheed Martin F-35 Lightning II joint strike fighter’s Block 4
radar. Upgrades will include a wide-area, high-resolution SAR
mode to the craft’s existing Northrup Grumman APG-81 radar.
The resulting upgrades, due in two stages in 2021 and 2023,
will yield “Big SAR” capability able to capture a significantly
larger ground area than what is possible with current systems.
This will aid in reconnaissance and targeting, with
GPGPU-driven processing allowing the system to handle the
much larger data load.

Within the military computing space, VPX (also known as VITA Group that designs and promotes VPX specifications, and
46, of which there are now many sub-specifications) has ADLINK maintains an ever-evolving roster of VPX and similar
emerged as a popular Eurocard/ backplane/chassis technology computing products well-suited to military signal analysis
for highly compact, dense computing platforms. VPX switched applications. ADLINK GPGPU-based products provide high
fabric backplanes provide high data throughput that allows for tech radar systems with digital signal processing and
simultaneous operations on large data sets. ADLINK is one of machine learning capabilities able to extract useful
the key members of the VME International Trade Association information from very high noise levels. These products
(VITA) VPX Working include:

Designed for rugged 3U enclosures, the VPX3010 processor blade features three CPU options:
the Intel® Xeon® D-1559 (12-core, 45W TDP), Intel® Xeon® D-1539 (8-core, 35W TDP),
and Intel® Pentium® D1519 (4-core, 25W TDP). The VPX3010 provides a power-efficient
processor base on which can be added a complementary GPU solution, such as discrete VPX
GPGPU blade or XMC GPGPU module through the XMC interface.

VPX Graphics Card

The 3U VPX GPGPU blade embeds dual-channel GDDR5 memory alongside the NVIDIA GPU.
Featuring hundreds of processing cores and CUDA compatibility, the VPX graphics card is
available in conduction-cooled (R) and air-cooled (A) variants.

For various reasons, including legacy investment support, some solutions may fare better
using CompactPCI rather than VPX. Like the VPX3010, the ADLINK cPCI-6940 processor
blade with Intel® Xeon® processor D-1500 and AMD Radeon™ E8860 GPU in 6U form factor
offers a high-performance foundation for radar processing systems.

Sonar Processing
Within the sonar sphere, digital signal processing can acoustic workloads comparing Intel® Core™ i7-4510U
extend to analysis of signals from towed and fixed acoustic (2 cores), NVIDIA GeForce 820M (96 cores, entry-level at
arrays, sonobuoys, torpedo guidance, and other systems. the time), and NVIDIA GeForce GTX 480 (480 cores, mid-level
Applications include the MK-48 torpedo, the Poseidon P-8, at the time) platforms. Not surprisingly, the GTX 480
and autonomous underwater vehicles (AUVs). As with radar, trounced its rivals.
GPGPU processing can perform the herculean task of
cutting through “salt and pepper noise” in a far faster, ADLINK’s broad range of GPGPU products provide sonar
efficient manner than CPU-only computation. This was receiver designers with multiple processing options based
proven by the University of Catania’s Placido Salvatore on SWaP considerations. A sampling of these includes:
Battiato when he benchmarked real-time image and

When SWaP priorities allow for more leniency in form factor, the 6U VPX form
factor delivers the most compute performance per rack unit for GPGPU applications.
ADLINK’s VPX6000 harnesses up to two Intel® Core™ i7-4700EQ (4-core, 47W TDP),
providing significant processing horsepower that remains within the bounds of
conduction-based cooling. However, as the Core™ i7 only provides integrated Intel
graphics, a companion GPU card is still required for a full GPGPU solution.

XMC Graphics Module

The XMC standard, also known as Switched Mezzanine Card, is one type of PCI
Mezzanine Card (PMC) defined by the VITA 42 standard. XMC specifies multiple
high-speed serial connection formats and offers an easy way to add modular,
cutting-edge I/O options to a platform without the heavy investments of a custom
solution. ADLINK’s XMC features surface-mounted GDDR5 memory and an NVIDIA
GPU. Various SKUs offer display output options as well as conduction or air cooling.

Image Processing
Image processing methods perform operations to enhance ADLINK’s HPERC (highly ruggedized, MIL-SPEC SFF
an image and/or extract useful information from it. standalone sub-systems) and GPGPU products, coupled
Applications include pattern of life analysis, surveillance, with NVIDIA’s CUDA graphics processing and the OpenGL
reconnaissance, target identification, and geolocation in API, provide solution engineers with powerful COTS
GPS-denied areas. For example, the U.S. Department of (commercial off-the-shelf) products with which they can
Defense collects tens of thousands of hours of aerial design the next generation of image processing systems.
surveillance video each month from Afghanistan and other Sample products include the aforementioned VPX GPGPU
theaters of operation. Turning what are often grainy, card, XMC GPGPU module, and others, including:
shaky videos taken by drones or satellites into clean image
streams suitable for analysis — especially real-time analysis
— poses a massive computational burden. According to
NVIDIA and MotionDSP, GPUs can process such video
seven times faster than CPU-only approaches.

Designed for operation in Extreme Rugged temperatures from -40°C to +85°C, the HPERC
meets VITA-75 specifications and weighs just over 3 kg. The two Intel® Core™ i7 options
(dual-core i7-3517UE or quad-core i7-3612QUE) dissipate heat through a VITA 75.22 spec
conduction coldplate. IO connectors are MIL-DTL-38999 standard, and internal expansion
busses include MXM, PCI Express Mini Card (Gen2), and PCI/104 Express Type 2.

MXM Graphics Module

This optional MXM module allows users to utilize NVIDIA’s GPGPU with GDDR5, making
HPERC a remarkably compact, highly versatile solution for fast image analysis, particular-
ly in remote and/or mobile settings.

Conclusion About ADLINK
Radar, sonar, and image processing may be some of the
Embedded Systems for Enabled Militaries
highest-growth niches for GPGPU applications within the
As defense strategies increasingly emphasize greater
military, but there are many more. Deep learning,
functionality and speed within ever smaller and more
artificial intelligence, gaming and simulation, cryptography,
rugged form factors, ADLINK stands uniquely positioned
and others can benefit from GPGPU assistance in similar
to help provide the information technology required by
measures. By pairing GPGPU advances with proven, reliable
tomorrow’s highly connected, data-driven forces. Founded
computing hardware from ADLINK, engineers can craft
in 1995, ADLINK designs and manufactures a broad range
military solutions able to deliver reliable, actionable
of products for embedded computing and automation
analysis results with unprecedented speed in a wide range
applications. Solutions span industrial motherboards,
of SWaP-minded solutions. Commanders have more data
computer-on-modules, extreme outdoor servers, network
to assess than ever before. Fortunately, the technologies
security, machine vision, and more. A Premier Member in
exist to turn that deluge into real-time intelligence
the Intel® Intelligent Systems Alliance with deep expertise
accessible in any theater, no matter how small or remote.
in IoT solutions, ADLINK continuously works with chip
manufacturers to bring new, better designs to customers.
ADLINK’s decades of experience and leadership in the
embedded market allow it to leverage highly competitive
cost strengths to industrial segment OEMs and systems
integrators around the world. At a time when effective
defense demands real-time results and exceptional
reliability, ADLINK delivers the building blocks needed to
surpass objectives and save lives.


ADLINK Technology, Inc. ADLINK Technology (China) Co., Ltd.

9F, No.166 Jian Yi Road, Zhonghe District 上海市浦东新区张江高科技园区芳春路300号 (201203)
New Taipei City 235, Taiwan 300 Fang Chun Rd., Zhangjiang Hi-Tech Park
新北市中和區建一路166號9樓 Pudong New Area, Shanghai, 201203 China
Tel: +886-2-8226-5877 Tel: +86-21-5132-8988
Fax: +886-2-8226-5717 Fax: +86-21-5192-3588
Email: Email:

Ampro ADLINK Technology, Inc. ADLINK Technology Beijing

5215 Hellyer Avenue, #110 San Jose, CA 95138, USA 北京市海淀区上地东路1号盈创动力大厦E座801室(100085)
Tel: +1-408-360-0200 Rm. 801, Power Creative E, No. 1 Shang Di East Rd.
Toll Free: +1-800-966-5200 (USA only) Beijing, 100085 China
Fax: +1-408-360-0222 Tel: +86-10-5885-8666
Email: Fax: +86-10-5885-8626

ADLINK Technology Singapore Pte. Ltd. ADLINK Technology Shenzhen

84 Genting Lane #07-02A, Cityneon Design Centre, 深圳市南山区科技园南区高新南七道数字技术园
Singapore 349584 A1栋2楼C区 (518057)
Tel: +65-6844-2261 2F, C Block, Bldg. A1, Cyber-Tech Zone, Gao Xin Ave. Sec. 7
Fax: +65-6844-2263 High-Tech Industrial Park S., Shenzhen, 518054 China
Email: Tel: +86-755-2643-4858
Fax: +86-755-2664-6353

ADLINK Technology Singapore Pte Ltd. ADLINK Technology GmbH

(Indian Liaison Office) (Mannheim) Hans-Thoma-Strasse 11, D-68163,
#50-56, First Floor, Spearhead Towers Mannheim, Germany
Margosa Main Road (between 16th/17th Cross) Tel: +49 621 43214-0
Malleswaram, Bangalore - 560 055, India Fax: +49 621 43214-30
Tel: +91-80-65605817, +91-80-42246107
Fax: +91-80-23464606 (Deggendorf) Ulrichsbergerstrasse 17, 94469
Email: Deggendorf, Germany
Tel: +49 (0) 991 290 94 – 10
ADLINK Technology Japan Corporation Tel: +49 (0) 991 290 94 - 29
〒101-0045 東京都千代田区神田鍛冶町3-7-4  Email:
Unizo Kanda Kaji-cho 3 Chome Bldg. 4F, 3-7-4 Kanda Kajicho ADLINK Technology, Inc.
Chiyoda-ku, Tokyo 101-0045, Japan
Tel: +81-3-4455-3722 (French Liaison Office)
Fax: +81-3-5209-6013 6 allée de Londres, Immeuble Ceylan 91940
Email: Les Ulis, France
Tel: +33 (0) 1 60 12 35 66
ADLINK Technology, Inc. Fax: +33 (0) 1 60 12 35 66
(Korean Liaison Office)

ADLINK Technology, Inc.

경기도 성남시 분당구 수내로46번길4 경동빌딩 2층
(수내동 4-4번지) (우) 13595
2F, Kyungdong B/D, 4 Sunae-ro 46beon-gil, Bundang-gu, (UK Liaison Office)
Seongnam-si, Gyeonggi-do, Korea, 13595
Toll Free: +82-80-800-0585 First Floor West Exeter House, Chichester Fields
Tel: +82-31-786-0585 Business Park Tangmere, West Sussex,
Fax: +82-31-786-0583 PO20 2FU, United Kingdom
Email: Tel: +44-1243-859677

ADLINK Technology, Inc.

(Israel Liaison Office)
27 Maskit St., Corex Building PO Box 12777
Herzliya 4673300, Israel
Tel: +972-54-632-5251
Fax: +972-77-208-0230


All Updated by May 11, 2018. ©2018 ADLINK Technology, Inc. All Rights Reserved.