Conference PaperPDF Available

Towards an Integrated GPU Accelerated SoC as a Flight Computer for Small Satellites

  • Turion Space Corp.
Towards an Integrated GPU Accelerated SoC as a Flight
Computer for Small Satellites
Caleb Adams
Department of Computer Science
University of Georgia
1510 Cedar St.
Athens, GA 30602
Allen Spain
Department of Electrical
and Computer Engineering
University of Georgia
314 Barnett Shoals Rd.
Athens, GA 30605
Jackson Parker
Department of Electrical Engineering
University of Georgia
1510 Cedar St.
Athens, GA 30602
Matthew Hevert
Department of Mechanical Engineering
University of Georgia
1510 Cedar St.
Athens, GA 30602
James Roach
Department of Computer Science
University of Georgia
1510 Cedar St.
Athens, GA 30602
Dr. David Cotten
Department of Geography
University of Georgia
210 Field St.
Athens, GA 30602
Abstract—Many small satellites are designed to utilize cutting
edge technology with the goal of rapidly advancing space based
capabilities. As a result, many components take advantage
of developments from the miniaturization of smartphone tech-
nology. Within the past 2 years, the UGA Small Satellite
Research Laboratory has extended this concept into embedded
GPUs for high-performance processing in LEO. Here we show-
case advances in our research of high-performance space-based
computation by integrating a traditional flight computer with
existing miniaturized GPU/SoC systems. Such a system paves
the way for many of NASA’s goals that require space based
AI, neural networks, computer vision, and high performance
computing. Our system fits a standard CubeSat PC/104+ form
factor, and implements many standard protocols such as I2C,
SPI, UART, and RS422. The system also has several GPIO pins,
2 USB-C ports, a micro USB port for flashing, an Ethernet port,
and a micro SD card slot for development. Additionally, the
system is designed to be modular, so that GPU accelerated SoCs
can be stacked to form a distributed system. For our primary
computer, which handles I/O and initializes processes on the
SoC, we choose to use the radiation tolerant Smart Fusion 2 SoC
with an ARM Cortex-M3 processor and a FPGA. In addition
to this primary computer, we use the Nvidia Tegra X2/X2i as
the GPU/SoC workhorse. The primary computer and the TX2i
are designed to share memory space with peripherals mounted
onto the board, so that no significant file transfer is required be-
tween the subsystems. Additionally, Nvidia’s Pascal architecture
enables GPU-CPU or GPU-GPU communication without PCIe,
enabling dense interconnected networks for monitoring and
computation. To address thermal concerns, we cap the TX2i’s
power draw at 7.5 Watts, provide recommendations for thermal
interface materials, and ensure that the primary computer only
enables the GPU/SoC when parallel computation is specifically
requested. Furthermore, radiation mitigation techniques are
explored with ECC, software mitigation techniques, and alu-
minized kapton sheets. In conclusion, this system is a step
towards a miniaturized high-performance flight computer well
suited for future computational demands.
1. INTRODUCTION................................... 1
2. FORM FACTOR ................................... 2
3. HARDWARE DESIGN .............................. 2
978-1-5386-6854-2/19/$31.00 c
2019 IEEE
4. ARCHITECTURE .................................. 3
6. CONCLUSION AND FUTURE WORK .............. 5
APPENDICES......................................... 6
ACKNOWLEDGMENTS ............................... 6
REFERENCES ........................................ 6
The University of Georgia’s Small Satellite Research Lab-
oratory (SSRL) is utilizing a Nvidia TX2i Graphics Pro-
cessing Unit (GPU) within the Multi-view Onboard Com-
putational Imager (MOCI) Satellite. Although the system
utilizes a GPU, an additional On Board Computer (OBC)
is still required for control and communication with core
avionics. The UGA SSRL has developed a board, the Core
GPU Interface (CORGI), that is capable of interfacing the
Nvidia TX1, Nvidia TX2, or Nvidia TX2i into a PC/104+
compliant CubeSat [1][2][3]. In this paper we discuss the
previously designed CORGI board and elaborate on lessons
learned, including how the CORGI is being used to combine
an OBC and an Nvidia TX2i into one board. We call this new
board, still undergoing development, the Accelerated Flight
Computer (AFC).
The GPU (Nvidia TX2/TX2i) being used is a complete Sys-
tem on a Chip (SoC), capable of running GNU/Linux on an
ARMv8, with a 256 core Pascal GPU. For an exhaustive list
of TX2/TX2i capabilities see the user guide [4] Currently the
TX2/TX2i utilizes CUDA 9.1 [4].
The Microsemi Smart Fusion 2 (SF2) SoC serves as the
central control node for the AFC. The SF2 SoC contains a
166MHz ARM Cortex-M3 Processor with embedded flash
and an FPGA[5]. The SoC has as static power draw of 7mW
and has Single Event Upset (SEU) protected and tolerant
eSRAM and DDR bridges [6]. The SF2’s FPGA has demon-
strated that it operates well in heavy ion and proton radiation
environments, though some mitigation is still required [7].
Additionally, past high-performance space processors have
used the SF2 SoC with success [8]. The architecture of the
SF2 SoC makes it ideal for space environments and its wealth
of documentation makes us confident that it will work well
with our design.
Figure 1.A previous iteration of the AFC concept,
known as the CORGI (Core GPU Interface).
Additionally, we seek to adhere to the IEEE Std 1156.4-1997
standard for Spaceborne Computer Modules. This standard
provides requirement levels for thermal performance, pres-
sure, shock, vibration, and radiation [9].
The CubeSat form factor is one of the primary considera-
tions in the design of the AFC. The form factor complies
with available deployers such as the Poly Picosatellite Or-
bital deployers (P-POD), and the JEM Small Satellite Or-
bital Deployer (J-SSOD), and is modular to accommodate
the payload and support I/O both spatially and functionally
[10][11]. The CORGI (see Figure 1) board was a design
benchmark which satisfied requirements, such as mission and
development requirements. However the design still required
optimization and additional features to provide a compre-
hensive accelerated heterogeneous computing platform[8].
CORGI’s successor, the AFC, has useful features such as bi-
directional logic shifters to convert the 1.8V (CMOS) logic
from the onboard Nvidia Jetson TX2i [4] to the Smart Fusion
2 3.3v (LVTTL) logic. Shifters are required for serial data
transfer, GPIO flipping, and IC enabling. The AFC will
maintain a PC/104+ form factor (90 x 96 mm)[3] with an
ergonomic cutout to accommodate the high speed LVDS ex-
pansion header (Samtec LSHM-120-04.0-L-DV-A-N-K-TR)
mounted beneath the TX2i module. Tight integration of
both CPU and GPU allows support of all aspects of satellite
performance, such as power, command and data handling,
attitude determination, payload etc. The Smart Fusion 2
ARM Cortex-M3 SoC (M2S150T-1FCG1152) is a low power
and low profile FCBGA (Flip Chip Ball grid Array) com-
bining program accelerators: FPGA, and CPU [8]. The
SF2, primary computer contains an ARM Cortex SoC and is
equipped with a flash based FPGA. The Nvidia TX2/TX2i’s
(50 mm x 87 mm) compact design provides high density
compute, and parallelized Pascal microarchitecture. A 400-
pin compatible board-to-board connector will interface with
the TX2i module via a 8x50 connector (SEAM-50-02.0-S-08-
2-A-K-TR). The Jetson TX2i was added to the AFC due to its
form factor and recent usage in high performance computing
(HPC) in industrial environments, long life, and ECC mem-
ory [4]. The AFC form factor maximizes compute, while
eliminating general purpose parts sold with COTS On Board
Computers (OBC). This has the added benefit of simplicity of
development, and reduced quiescent current draw.
Table 1.Processors
Smart Fusion2:
ARM Cortex-M3 SoC (M2S150T-1FCG1152)
NVIDIA Jetson TX2:
-2x Denver ARM Cortex-A57
-256-core Pascal GPU
Table 2.Memory
SF2: 4x256 DDR3 Memory Bank
TX2: 8GB LPDDR4 and 32 GB ECC support
-2 Gb (Cypress CYRS16B256)
-1 Gb SPI Flash on SPI0 via FPGA MSS
-1Gb Flash on SPI1 via FPGA fabric
Table 3.IO Interfaces
PC/104+, 12C, SPI, GPIO
Expansion header (QSPI)
2x RJ-45
2x USB 3.0 Type C
Micro USB (FTDI)
As shown in Figure 3, H1 and H2 are the main PC/104+
CubeSat headers. These route peripheral components onto
the bus, as well as routing the PC/104+. The H1 and H2
headers on the AFC also act as pass-throughs, so there is no
break in communication or power. The Smart Fusion 2 shares
all primary communication lines and acts as the control node
for the command and data handling (CDH) subsystem of the
satellite. In addition to two RJ45 Ethernet receptacles the
AFC board also makes a QSPI expansion header available for
added support and 2x USB Type C for debugging the TX2i.
The SmartFusion 2 advanced development kit [12] was used
for the prototyping of the integrated CPU-GPU board. The
SmartFusion 2 advanced development kit contains a wide
range of headers (SPI, I2C, GPIO, UART, etc.) [12]. The
versatility of this development kit makes it ideal for prototyp-
The Nvidia TX1/TX2 development kit includes a developer
board that is compatible with the Nvidia TX2i [4]. The
developer board, connected to the Nvidia TX2/TX2i, can be
interfaced with the Smart Fusion 2 advanced development
kit via 10/100/1000 PHY Ethernet connection, SPI, LVTTL
UART, I2C, PCIe, or GPIO[4][12].
Figure 2.AFC 3D design back
Figure 3.AFC 3D design front
Hardware Architecture
With ubiquitous computing and 10/100/1000 Ethernet
Switching, Ethernet provides wired communication for the
satellite bulk data transfer between the SF2 and the one
or more TX2i. This design contains carrier level Ethernet
whose data link and physical layer protocol are described
by the IEEE 802.3. [13] This design utilizes a node switch
which acts as a physical medium.The SF2, being the pri-
mary distributer of the workload, design contains SEU Pro-
tected/Tolerant Memories i.e: eSRAMs, DDR3. Its robust
nature played a considerable roll in its selection as the pri-
mary processor on the accelerated flight computer. In the
event of one or all of the TX2is destruction, the SF2 will be
able to perform I/O monitoring and control duties without a
co-compute device.
This option, although functional, will diminish payload ca-
pabilities. In response, the system-on-chip (SoC) field pro-
grammable gate array (FPGA) microcontroller subsystem
(MSS) includes a watchdog which provides a system which
needs no operating system to detect and respond to lockups
caused by heavy ion bombardment [12]. The SF2 will
utilize 4 x 256 DDR3 (MT41K256M8DA-125 IT:K) external
memory banks providing volatile memory. In addition, the
TX2i modules and SF2 share RAD NOR SPI flash memory
(Cypress CYRS16B256). The SF2 will allow mediated
read/write access with the purpose of sharing mutually rele-
vant files. This functionality is defined in hardware via MUX
whose selection is determine by GPIO pins from SF2 SoC,
active high.
Software Architecture
This section details how software will make use of the hard-
ware system presented above in order to provide a modular
fault tolerant accelerated computing platform. The Smart
Fusion 2 will act as the main control node for the whole
system and will make use of four main interfaces to interact
with the TX2is. These interfaces will be a Gb Ethernet
local network, GPIO, SPI, and shared flash memory that the
Smart Fusion 2 and TX2is will access via multiplexed SPI
connections. The ethernet line will be used by the Smart
Fusion 2 to distribute computation tasks and data to the
TX2i’s. These ethernet communications will take the form
of multicast messages, as each TX2i will need to receive and
direct messages for communication to individual TX2is. The
computation tasks to be distributed to each TX2i will take the
form of CUDA code. The TX2is will make use of the ethernet
line or GPIO to kick the watchdogs on the Smart Fusion 2
FPGA and to log their progress or telemetry. The Smart
Fusion 2 will use the SPI accessed shared flash memory as
its primary persistent memory where it will store tasks and
data to distribute to the TX2is. The TX2is will make use of
the shared flash memory by writing computational results to
System Hardware Mitigation and Shielding
For radiation and thermal mitigation in Low Earth Orbit
(LEO) we recommend utilizing the Dunmore Aerospace
“Satkit” which contains the standard STARcrest MLI mate-
rials cut into manageable sizes for small satellite developers.
This allows one to develop their own thermal protection
blankets according to various mission requirements. The kit
includes an outer layer material, inner layer material, first
surface tape, and clear polyimide tape[14]. This kit is optimal
for small satellite systems as it is small, cheap, and easily
customizable. More specifications of the materials are listed
below (See Figure 4) and values were calculated from the
Solar Radiation, Earth’s IR radiation, and Albedo Radiation
equations [15].
q=Gsscos (1)
eIRF e (2)
q=Gs(AF )sFecos (3)
Figure 4.The Hardware Design of the Cubesat GPU/SoC-CPU system
The heat flux due to the LEO sources using equations 1, 2,
and 3 was calculated with the help of data from literature [15].
Values have an expected error of ±0.4. Values for the solar
radiation were determined using the mean values provided by
data from the World Radiation Center in Davos Switzerland
The necessary thickness of the radiation shield was calculated
in part using values provided by Dunmore. The primary
source of heat flux is solar, these values vary on an annual
basis due to the Sun’s elliptical orbit. Meaning that the
maximum and minimum amount of flux expected from the
Sun would range between 1322 and 1414w/m2. The target
ambient temperature inside the CubeSat is 293.15 K (20
degrees C). This is calculated with the following formula:
Where Qis the heat flux into the system, Kis the thermal
conductivity of the material, A is the area in meters squared,
and Lis the thickness of the radiation shielding. Assum-
ing a thermal conductivity of 0.014 and a Tof 101K.
The expected required thickness of the radiation shielding is
1.03438 ·107m2. Radiation from Free Molecular Heating
(FMH) was determined to be negligible for the stage in which
the CubeSat would be launched from the ISS. FMH is almost
exclusively encountered during launch ascent just after the
booster’s payload fairing is ejected.
Radiation Mitigation
The radiation shielding thickness (See Equation 4) is also
driven by the 1997 IEEE Standard for Environmental Specifi-
cations for Spaceborne Computer Modules [9]. We design for
level I radiation in preperation for the LEO environment. An
additional benefit of the Dunmore Aerospace “Satkit” is that
it meets this standard while accruing minimal mass gains.
Proper shielding does not have to incur heavy mass gains,
Table 4.Recommended Dunmore Aerospace Satkit Part
VDA / 200 GA Polyimide
Tensile Strength 24000 psi
Elongation 50%
Thickness 50.8 micron
Density 1.42 g/cc
Yield 13.8 m·m/kg
Weight/Area 72 g/m2
Operating Temp -250 - 400 C
Metalization 99.99 % pure aluminium
VDA / 200 GA Polyimide
Tensile Strength 26000 psi
Elongation 110 %
Thickness 6.35 micron
Density 1.39 g/cc
Yield 124.9 m·m/kg
Weight/Area 8 g/m2
Operating Temp -250 - 150 C
Metalization 99.99 % pure aluminium
VDA / 25 GA PET / VDA, Embossed & Perforated
Thickness 76.2 microns
Yield 10.4 m·m/kg
Weight/Area 96 g/m2
Operating Temp -40 - 220 C
Metalization 99.99 % pure aluminium
100 GA Polyimide / 966 PSA
Thickness 76.2 microns
Yield 10.4 m·m/kg
Weight/Area 96 g/m2
Operating Temp -40 - 220 C
Metalization 99.99 % pure aluminium
and in fact shielding that is too thick can increase the effects
of some kinds of radiation events. This is due to the higher
levels of secondary particles created when a high-energy
GCR particle impacts a thick shield [17]. If necessary, a
properly designed shield may act as a heat sink due without
exceeding mass limitations. This reduction of a device’s
operating temperature can greatly reduce the risks posed to
that device by radiation [17].
When working with Commercial Off The Shelf (COTS) elec-
tronic components for use in the space environment, some
of the problems that must be addressed are data integrity
performance and accuracy in high radiation environments. In
our design, the TX2i, one of the Tegra SoCs, is one of the
more vulnerable parts with respect to this, due to its highly
dense hardware design[18]. As stated above, much can be
done at a hardware level but often times additional software
mitigation is needed.
One of the major problems radiation presents for software is
its effect on memory. Overtime as bit flips [15] [9] aggregate,
they will corrupt information and make some systems unus-
able. To address these issues, in the AFC we plan on modify-
ing the file system and bootloader. We will triplicate the file
system on the TX2i so that any part of it which is corrupted
will be able to use the principle of triple modular redundancy
(TMR) to automatically correct damage. In addition to this,
we will use the U-Boot bootloader software to modify the
bootloading process with TMR [19]. The bootloader will
store 3 copies of the operating system (OS) image for the
TX2i and a hash for each image. At boot time the bootloader
will recalculate the hash for each OS image it attempts to
load and compare the calculated hash against the stored hash
to determine if the OS has been corrupted. If corruption is
detected then boot loader will attempt to load the next OS
image. If all OS images are determined to be corrupted, the
bootloader shall attempt to construct an uncorrupted image
by bit voting between the corrupted images.
Our design uses the SF2 FPGA as a trusted control node
due test results showing high levels of radiation tolerance [7].
Thus, we plan to have the Smart Fusion 2 FPGA operate as a
watchdog for each TX2/TX2i. If the Smart Fusion 2 detects
that one of the TX2/TX2i’s has anomalous power draw levels
(over 7.5 Watts), it will send a command over ethernet (via
the SF2 SoC) commanding the TX2i to reduce GPU usage
until the power level stabilizes.
Thermal Mitigation
Simulations have shown, given a large enough mounting
structure, that the Nvidia TX2 is capable of dissipating heat
effectively. To achieve this goal it is imperative that the TX2’s
Thermal Transfer Plate (TTP) is adequately interfaced into
the heat dissipating mass[1]. To interface the TTP with the
heat dissipating mass we highly recommend the use of a low
thickness and high conductance thermal interface material
(TIM) that adheres to the NASA standards for collected
volatile condensible materials (0.1% CVCM) and total
mass loss (1% TML) [20]. Previous findings have sug-
gested that the Carbice Space TIM is ideal for these purposes
due to its low thickness (0.065mm) and high conductance
2K1) [1].
It is important to note that all thermal simulations are run
with a heat load according to subsystem power draws and
assume 0% efficiency for a worst case scenario. The steady
state thermal simulation clearly shows that the GPU’s core
temperature has been lowered with the addition of the TIM.
The overall maximum has been brought down from around
160C, to under 50C. While this model did not include
the entire spacecraft structure as a conduction medium, this
would only serve to decrease the temperature further, as
Figure 5.The Nvidia TX2 daughter-board with the
Parker Series SoC Exposed. No TIM is used in this
Figure 6.The Nvidia TX2 daughter-board with the
Parker Series SoC Exposed. TIM is used in this
conduction is a much faster heat transfer mechanism than
Figure 7.The Nvidia TX2 with TTP integrated and TIM
However, some caution must be used with these results. The
thermal environment of the satellite is inherently transient, so
the ambient temperature assigned to the model here is likely
inaccurate. This also means a “steady state” analysis might
not necessarily be appropriate. However, the purpose of
this analysis is not to provide conclusive thermal information
about the AFC, rather to serve as a data point for design,
and to show that under extremely approximate conditions, the
AFC will be able to operate in orbit.
The UGA SSRL’s CORGI design, the starting point for de-
signing the AFC, has made the usage of the Nvidia TX2/TX2i
more feasible in LEO. The recommendations provided in the
Figure 8.The Nvidia TX2 with TTP integrated, no TIM
is used.
paper regarding radiation and thermal mitigation provide a
starting point for those who wish to utilize the Nvidia TX1,
TX2, or TX2i in LEO. Additionally, when designing a fully
integrated system with the Nvidia TX2/TX2i, the use of a
SmartFusion2 SoC FPGA is highly recommended. It is well
understood, well documented, radiation tolerant, and operates
on low power. The combination of the SmartFusion2 with the
Nvidia TX2/TX2i has the potential to meet many of NASA’s
goals that require space based AI, neural networks, computer
vision, and high performance computing[21].
Future work will entail testing of the aforementioned compo-
nents. The system shall be able to halt and restart actions
on the subsystem and utilize Paralleled Thread Execution
assembly to checkpoint GPU operations. Additional methods
for thermal dissipation will be simulated and those mentioned
above will be tested in our thermal vacuum chamber. Radia-
tion testing (heavy ion and proton) will be performed during
heavy computational load. The system will be used to form
a modular computational cluster and distributed computation
will be further tested in our vacuum chamber while the system
is under heavy load.
The resources used in this paper can be requested at any point
by contacting the authors of this paper.
The authors would like to thank the Georgia Space Grant
Consortium for funding these GPU research projects and
the Air Force Research Laboratory’s University Nanosat
Program for giving us tremendous opportunities and for
funding the projects that led us to this point. Additionally
the authors would like to thank Erick Gavilanes for making
the initial design for the CORGI in 2017 and getting us to
a point where we could perform iterations and follow up
research. Thanks to Casper Versteeg for providing significant
knowledge and results relating thermal mitigation. Nir Patel,
Nicholas Heavner, and Michael Buzzy did great work by
contributing to the thermal models. A big thanks Justin
Heimerl, for supplying CAD and working on electrical design
and to Sydney Whilden and Adam King for doing initial
radiation research and laying the foundations for this paper
in that respect. A special thanks to Dr. Deepak Mishra for
helping support the University of Georgia’s Small Satellite
Research Laboratory from the beginning. Another thank you
to Roger Hunter.
[1] C. Versteeg, “Thermal management and design of high
heat small satellite payloads,” 32nd Annual AIAA/USU
Conference on Small Satellites, 2018.
[2] C. Adams, A near real time space based computer vi-
sion system for accurate terrain mapping,” 32nd Annual
AIAA/USU Conference on Small Satellites, 2018.
[3] (2003) Pc/104-plus specification. [Online]. Avail-
[4] (2017) Jetson tx2 developer kit user guide.
[Online]. Available:
[5] (2015) Arm cortex-m3 processor tech-
nical reference manual. [Online]. Avail-
doc.ddi0337e/DDI0337E cortex m3 r1p1 trm.pdf
[6] (2018) Product brief smartfusion2 soc fpga. [Online].
doc view/134215-ug0557-smartfusion2- soc-fpga-
[7] (2014) Igloo2 and smartfusion2 65nm commercial
flash fpgas interim summary of radiation test results.
[Online]. Available:
portal/doc view/134103-igloo2-and- smartfusion2-
[8] F. Bruhn, “Introducing radiation tolerant heterogeneous
computers for small satellites,” 32nd Annual AIAA/USU
Conference on Small Satellites, 2015.
[9] “Ieee standard for environmental specifications for
spaceborne computer modules,” IEEE Std 1156.4-1997,
1997. [Online]. Available:
[10] (2016) 6u cubesat design specification. [Online].
Available: 1/
[11] “Nanoracks cubesat deployer (nrcsd) interface
definition document (idd). nr-nrcsd-s0003, 2018.
[Online]. Available:
[12] (2016) Smartfusion2 soc fpga advanced development
kit user guide. [Online]. Available: www.microsemi.
[13] (2008) Ethernet theory of operation. [Online]. Avail-
[14] Dunmore aerospace satkit. [Online]. Available: www.
[15] G. D. Badhwar, “The radiation environment in low-
earth orbit,” Radiation Research, vol. 148, no. 5,
pp. S3–S10, 1997. [Online]. Available:
[16] “World radiation data centre, 2018. [Online].
[17] R. H. Maurer, M. E. Fraeman, M. N. Martin, and
D. R. Roth, “Harsh environments: Space radiation en-
vironment, effects, and mitigation,Johns Hopkins APL
Technical Digest, vol. 28, no. 1, pp. 17–29, 2008.
[18] E. Wyrwas, “Body of knowledge for graphics
processing units (gpus) nepp-bok-2018,” 2018.
[Online]. Available:
atoms/files/2017-8-1 stip final-508ed.pdf
[19] “The denx u-boot and linux guide (dulg) for
canyonlands,” 2018. [Online]. Available: www.denx.
[20] R. R. J. Roe, “Standard materials and processes require-
ments for spacecraft,” Technical Report NASA-STD-
6016A, November 2016.
[21] (2017) Nasa strategic technology investment plan.
[Online]. Available:
atoms/files/2017-8-1 stip final-508ed.pdf
Caleb Adams Caleb Adams is a Grad-
uate student in Computer Science and
co-founder of the UGA Small Satel-
lite Research Laboratory, which now
has 2 satellite missions. Caleb’s re-
search mostly focuses on high perfor-
mance computation systems in space en-
vironments for AI and Computer Vision.
He has also worked on audio telemetry
systems for Orion at NASA, was also a
google glass beta tester, and a TEDxUGA speaker.
Allen Spain Allen Spain is currently
a M.S. Student in the School of Elec-
trical and Computer Engineering at the
University of Georgia. He specializes
in electronic circuit design, and is the
primary designer of Electronics Inter-
face Boards for the UGA Small Satellite
Laboratory. He is currently a Research
Assistant at the University of Georgia
focusing on Biomimetic Photonic algo-
Jackson Parker Jackson Parker re-
ceived his B.S. in biological engineering
in spring of 2018 from the University of
Georgia, where he is currently a Mas-
ters student in Electrical and Computer
Engineering. Current research activities
and interests include accelerated com-
putation, machine learning, algorithmic
development, and computer vision. He
plans on furthering his knowledge in
those areas of interest while working on his M.S. and then
pursue a PhD in computer science.
Matthew Hevert Matthew Hevert is
currently pursuing a Bachelors of Sci-
ence in Mechanical Engineering at the
University of Georgia with an antici-
pated graduation in 2020. He has been
working at the Small Satellite Research
Laboratory performing work in optics,
thermodynamics, and mechanical radia-
tion mitigation. He plans on continuing
his education after graduation in optical
sciences and engineering where he one days hopes to help
design the next generation of space-based telescopes.
James Roach James Roach is an un-
dergraduate computer science student
graduating from the University of Geor-
gia in May 2019. He has been working
at the small satellite research laboratory
since fall 2015. He has also worked on
small satellites for Space Dynamics Re-
search Laboratory. Current research in-
terests include accelerated computation,
the effect of radiation on electronics, and
fault tolerance in cubesat platforms.
David Cotten David L. Cotten, Ph.D.,
received a B.S in Physics from Louisiana
State University in 2005. He currently
serves as the manager and Co-PI of the
UGA’s Small Satellite Research Labora-
tory and is an Assistant Research Sci-
entist at the Center for Geospatial Re-
search in the Geography Department at
UGA. He graduated from UGA in 2011
with his doctorate in Physics and As-
tronomy. His research as a Post-Doctoral Associate focused
on surface-atmosphere exchange, and he is currently using
remote sensing (multispectral/hyper spectral sensors) and
micrometeorology techniques to quantify carbon storage in
wetland regions at both the local and regional scales. Dr.
Cotten is also using unmanned aerial vehicles, air photos,
and satellite imagery to build 3D models of terrestrial objects
using photogrammetric Structure from Motion.
... When radiation, thermal and vibration resilience are of utmost importance, space-grade FPGAs are used instead of their COTS counterparts to achieve increased reliability. The literature also includes numerous works with space avionics co-processing architectures that include FPGAs [332][333][334]. In terms of algorithms, besides accelerating DSP, the FPGAs are used for implementing data transcoding for instruments/sensors (e.g., via SpaceWire/SpaceFibre [335]) and data compression [336,337]. ...
... More details about their use in past, present and future missions can be found in our publication in [320]. At research level, the space-grade FPGAs are being evaluated for the implementation of compute-intensive functions and novel space applications, while they are also examined as part of co-processing architectures [332][333][334]. Next, we present some representative research works involving space-grade FPGAs. ...
... The use of COTS components in Low Earth Orbit (LEO) missions and CubeSats relies on the partial shielding provided by Earth's magnetosphere and/or the short mission lifetime, which limit the damage or unavailability of electronics due to radiation. In this context, FPGAs [324][325][326]328], GPUs [332,333,[371][372][373] and VPUs [374][375][376][377][378] are evaluated as accelerators, while they are also subjected to radiation tests, such as the Myriad 2 VPU [379]. A second challenge for the space industry is the wider adoption of AI, which is currently limited to offline/ground data processing and not on-board processing, mostly due to insufficient computational power and increased memory footprint, as well as qualification issues when deployed in orbit [379]. ...
Full-text available
The computing industry is forced to find alternative design approaches and computing platforms to sustain increased power efficiency, while providing sufficient performance. Among the examined solutions, Approximate Computing, Hardware Acceleration, and Heterogeneous Computing have gained great momentum. In this Dissertation, we introduce design solutions and methodologies, built on top of the preceding computing paradigms, for the development of energy-efficient DSP and AI accelerators. In particular, we adopt the promising paradigm of Approximate Computing and apply new approximation techniques in the design of arithmetic circuits. The proposed arithmetic approximation techniques involve bit-level optimizations, inexact operand encodings, and skipping of computations, while they are applied in both fixed- and floating-point arithmetic. We also conduct an extensive exploration on combinations among the approximation techniques and propose a low-overhead scheme for seamlessly adjusting the approximation degree of our circuits at runtime. Based on our methodology, these arithmetic approximation techniques are then combined with hardware design techniques to implement approximate ASIC- and FPGA-based DSP and AI accelerators. Moreover, we propose methodologies for the efficient mapping of DSP/AI kernels on distinctive embedded devices, i.e., the space-grade FPGAs and the heterogeneous VPUs. On the one hand, we cope with the decreased flexibility of the space-grade technology and the technical challenges that arise in new FPGA tools. On the other hand, we unlock the full potential of heterogeneity by exploiting all the diverse processors and memories. Based on our methodology, we efficiently map computer vision algorithms onto the radiation-hardened NanoXplore's FPGAs and accelerate DSP & CNN kernels on Intel's Myriad VPUs.
... In the publication titled NVIDIA GVDB Voxels, this architecture was described in detail. [88] The system is capable of managing tens of millions of particles within a simulation region that is almost unlimited in size, it offers novel approaches to the establishment of a parallel, sparse grid hierarchy, and it offers fast incremental updates for transferring particles to the GPU. These are just a few of the capabilities that the system possesses. ...
... speedup over the state-of-the-art algorithms. [87] large chemical and biological systems new heterogeneous CPU+GPU high computational performance [88] Space for Small Satellites as a Flight Computer to meet many of NASA's goals that require space based AI [89] Fast Fluid Simulations FLIP technique Faster than running on the CPU. [90] bees swarm optimization GBSO-Miner up to 800 times faster than an optimized CPU Implementation [91] Space application low power embedded GPU Reduce data corruption from up to 46% to 2%, execution time overhead of 130%. ...
Full-text available
There have been a growing number of academics and developers who have appeared in recent years, and they have spent a significant amount of their time and attention to researching GPU, as well as the numerous sectors in which it plays an essential and successful function. This is in part due to the immense progress that the world is seeing on a daily basis, which includes a variety of disciplines of study such as medical, engineering, space, and precise sciences that demand high-resolution visual representation and vast graphic processing power. This article acts as a review and provides a condensed description of the development, applications, problems, and related research into GPU across a wide number of various industries. It should not come as a surprise, given all of these benefits, that cloud computing has become the standard practice in its business in such a short length of time. The movement of corporate operations to the cloud, which is being done by a growing number of companies, is being done in an effort to minimize the amount of time spent on infrastructure maintenance. This migration is taking place for the following reasons: As a consequence of this, preserving a cloud environment in such a manner that it continues to operate normally is a tremendously challenging endeavor. If you want to increase cloud performance and minimize the amount of administrative stress you suffer, you need a reliable cloud monitoring solution. The cloud monitoring service is an investment that is well worth making since there is a possibility that it will boost performance and minimize the amount of time that is spent managing. One of the most important roles of cloud monitoring is the management of Quality of Service (QoS) metrics for cloud-hosted, virtualized, and physical services and applications.
... Indeed, once the target CPU has been determined, the OpenCV optimised for OpenVino can handle the rest of the setup. b) Jetson TX2: The NVIDIA Jetson TX2 [28] is a high-performance AI on-the-edge devices with multiple GPU and CPU cores, capable of handling loads for an efficient implementation of CNN, and it has been already considered for space applications [29]. The TX2 device includes the Tegra X2 system-on-chip (SoC) design processor. ...
Conference Paper
Full-text available
This paper investigates the opportunity to use artificial intelligence methodologies and edge computing approaches for wildfire detection directly from satellite platforms. The test case for our study is PRISMA (Precursore IperSpettrale della Missione Applicativa-Hyperspectral Precursor of the Application Mission), the Italian hyperspectral satellite launched in 2019 by the Italian Space Agency. This mission provides hyperspectral (HS) images in the spectral range of [0.4, 2.5] µm and an average spectral resolution less than 10 nm. This work reports new results related to the Australian bushfires happened in December 2019 in New South Wales, captured by PRISMA on December 27, 2019. Starting from a one-dimensional convolutional neural network (CNN) discussed in previous authors' works to perform multiclass classification, this paper primarily deals with the opportunity to use hardware accelerators, namely the Intel Movidius Myriad 2, the Nvidia Jetson TX2, and the Nvidia Jetson Nano, to consider the on-the-edge implementation of the CNN. This study is in line with the current impulse to improve on-board computing capabilities and platform autonomy, setting some of the elements for future satellites or constellations focusing on specific remote sensing tasks to provide real-time reliable early warnings.
... The NVIDIA Jetson TX2 is one of the available high-performance AI on-the-edge devices with multiple GPU and CPU cores, capable of handling loads for an efficient implementation of CNN, and it has been already considered for space applications. 5 The TX2 device includes the Tegra X2 system-on-chip (SoC) design processor. The Tegra X2 incorporates a quad-core 2.0-GHz 64bit ARMv8 A57 processor, a dual-core 2.0-GHz superscalar ARMv8 Denver processor, and an integrated Pascal GPU. ...
Conference Paper
Full-text available
This work presents MONSTER, a Moon optical navigation robotic facility on simulated terrain, i.e., an experimental facility that can be used to simulate lunar navigation problems. The facility consists of a 3D Cartesian manipulator that, once fully operative, will be equipped with a spherical joint allowing to simulate both attitude and orbital dynamics. All the experiments that have been performed so far and are planned to be performed in the future are based on innovative and disruptive approaches using artificial intelligence (AI) algorithms. Indeed, a crater detection algorithm based on a fully convolutional neural network has been implemented, and a reinforcement learning approach is under development for prescribing the control policy of the simulated system. MONSTER enables hardware-in-the-loop simulations of landers and spacecrafts using AI hardware accelerators such a graphics processing unit (GPU), visual processing unit (VPU) and field programmable gate arrays (FPGA).
... A co-processing FPGA & VPU architecture is evaluated in [3], where the VPU is used for accelerating VBN pipelines with limited power consumption. The hybrid architecture of [6] bases on the SmartFusion2 SoC FPGA and the Tegra X2/X2i GPU (main accelerator). In [7], an heterogeneous architecture is proposed, consisting of a SoC FPGA for SpaceWire I/O transcoding, the AMD SoC (CPU & GPU) for acceleration, and optionally a VPU for AI deployment. ...
Conference Paper
Full-text available
The advent of computationally demanding algorithms and high data rate instruments in new space applications pushes the space industry to explore disruptive solutions for on-board data processing. We examine heterogeneous computing architectures involving high-performance and low-power commercial SoCs. The current paper implements an FPGA with VPU co-processing architecture utilizing the CIF & LCD interfaces for I/O data transfers. A Kintex FPGA serves as our framing processor and heritage accelerator, while we offload novel DSP/AI functions to a Myriad2 VPU. We prototype our architecture in the lab to evaluate the interfaces, the FPGA resource utilization, the VPU computational throughput, as well as the entire data handling system's performance, via custom benchmarking.
Graphics Processing Unit (GPU) devices and their associated software programming languages and frameworks can deliver the computing performance required to facilitate the development of next-generation high-performance safety-critical systems such as autonomous driving systems. However, the integration of complex, parallel and computationally demanding software functions with different safety-criticality levels on GPU devices with shared hardware resources contributes to several safety certification challenges. This survey categorizes and provides an overview of research contributions that address GPU devices’ random hardware failures, systematic failures and independence of execution.
China plans to develop the next generation dark matter particle explorer satellite, referred to as the Very Large Area Space Telescope (VLAST). As an essential step in this process, the prototype design of detectors and electronics for the VLAST is currently underway. The nuclide detector is a core detector in the VLAST. It mainly measures nuclides’ charges and distinguishes high-energy gamma rays and electrons. This paper will discuss the prototype readout electronics for the VLAST’s nuclide detector, which accurately measures the charge signal of the photomultiplier tubes using the VATA160 application-specific integrated circuit chip; furthermore, we consider a series of critical problems, including radiation-hardening and environment monitoring. The test results show that the system exhibits stable operation, good performance, and good technical indicators. Furthermore, each electronic channel achieves a dynamic range of 0–12.5 pC, the random noise level exceeds 1.6 fC, and the integral nonlinearity exceeds 0.35%.
Space-based moving targets tracking and observation facilitates target recognition and analysis of target characteristics, but the ability of satellite attitude tracking control needs to be improved, especially considering the energy optimization for long-time tracking. An attitude tracking control method combining deep reinforcement learning and predefined-time stability is proposed, which not only improves the autonomous decision-making ability but also ensures the reliability of the satellite attitude control. The long short-term memory network is integrated into the twin delayed deep deterministic policy gradient algorithm to learn the moving state of the target from its image positions as the input to generate the desired attitude in real time, and energy optimization is considered in the design of the reward function. Then an adaptive backstepping controller is designed to achieve predefined-time stability in the presence of the external disturbance and uncertain inertia properties, which ensures that the satellite attitude is controlled to the desired value within a predefined decision period. Finally, a simulation system of moving target tracking is established, and the results indicate that our approach is superior in terms of tracking ability and energy consumption.
Conference Paper
Full-text available
The Multiview Onboard Computational Imager (MOCI) is a 3U cube satellite designed to convert high resolution imagery, 4K images at 8m Ground Sample Distance (GSD), into useful end data products in near real time. The primary data products that MOCI seeks to provide are a 3D terrain models of the surface of Earth that can be directly compared to the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) v3 global Digital Elevation Model (DEM). MOCI utilizes a Nvidia TX2 Graphic Processing Unit (GPU)/System on a Chip (SoC) to perform the complex calculations required for such a task. The reconstruction problem, which MOCI can solve, contains many complex computer vision subroutines that can be used in less complicated computer vision pipelines.
Full-text available
adiation effects in solid-state microelectronics can be split into two general categories: cumulative effects and single-event effects (SEEs). Cumulative effects produce gradual changes in the operational parameters of the devices, whereas ior in circuits. The space radiation environment provides a multitude of trapped, solar, and cosmic ray charged particles that cause such effects, interfere with space- system operation, and, in some cases, threaten the survival of such space systems. This article will describe these effects and how their impact may be mitigated in
The Lunar Reconnaissance Orbiter (LRO) and Chandrayaan-1 both carry scientific investigations designed to measure the energetic particle fluxes and radiation dose in low lunar orbit. Mission operations overlapped for more than two months thus providing simultaneous measurements in low, polar lunar orbit. This was a time period of low solar activity and therefore the energetic radiation exposure was confined to the galactic cosmic rays (GCR) and albedo from the lunar surface created by interactions of the GCR with the lunar surface. The sensor basics and a survey of the observations will be presented.
The radiation environment in low-Earth orbit is a complex mixture of galactic cosmic radiation, particles of trapped belts and secondary particles generated in both the spacecraft and Earth's atmosphere. Infrequently, solar energetic particles are injected into the Earth's magnetosphere and can penetrate into low-Earth orbiting spacecraft. In this paper, the sources of charged-particle radiation that contribute significantly to radiation exposure on manned spacecraft are reviewed briefly, and estimates of expected dose rate for the upcoming International Space Station that are based on measurements made on the Russian Mir orbital station are provided.
Thermal management and design of high heat small satellite payloads
  • versteeg
C. Versteeg, "Thermal management and design of high heat small satellite payloads," 32nd Annual AIAA/USU Conference on Small Satellites, 2018.
Body of knowledge for graphics processing units (gpus) nepp-bok-2018
  • E Wyrwas
E. Wyrwas, "Body of knowledge for graphics processing units (gpus) nepp-bok-2018," 2018. [Online]. Available: atoms/files/2017-8-1 stip final-508ed.pdf
Standard materials and processes requirements for spacecraft
  • R R J Roe
R. R. J. Roe, "Standard materials and processes requirements for spacecraft," Technical Report NASA-STD-6016A, November 2016.