A Study of the SEU Performance of InP and SiGe Shift Registers

Abstract

Shift registers fabricated using InP and SiGe technology are tested for SEU performance when irradiated with protons and heavy ions. The results are compared to several different models which predict proton cross section from heavy-ion data.

Full text

Page 1

1140IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 52, NO. 4, AUGUST 2005
A Study of the SEU Performance
of InP and SiGe Shift Registers
D. L. Hansen, P. W. Marshall, R. Lopez-Aguado, K. Jobe, M. A. Carts, C. J. Marshall, P. Chu, and S. F. Meyer
Abstract—Shift registers fabricated using InP and SiGe tech-
nology are tested for SEU performance when irradiated with pro-
tons and heavy ions. The results are compared to several different
models which predict proton cross section from heavy-ion data.
Index Terms—Heavy ion, indium phosphide, proton, silicon ger-
manium, single event upset.
I. INTRODUCTION
S
enough to operate in an environment significantly more hos-
tile than conditions typically found on earth. In addition to
performance degradation from long-term exposure to radia-
tion, electronics are also subject to single event upsets (SEU),
initiated when a single cosmic particle incident on the circuit
interacts with a transistor to corrupt data. Previous studies
[1]–[4] have shown that SEU rates increase with frequency.
Consequently, the importance of SEU has grown as microelec-
tronic devices have become faster.
Within the earth’s magnetic field, SEU are caused primarily
by reactions with protons. The peak flux behind typical satel-
lite enclosures is at about 1.5 earth radii (
cantly by
In contrast, heavy ions are almost nonexistent in
low-inclination, low earth-orbits. Instead, they pose the greatest
threat at geosynchronous orbits (
traversesthesensitivevolumeofasemiconductordepositingen-
ergy this excites electrons beyond the ionization threshold and
creates electron-hole pairs. If the charge is sufficient, one or
more bits in the clock or data stream will be altered and an upset
occurs. Proton induced SEU often involve a second indirect
mechanism. Incident protons interact with nuclei in the semi-
conductor to create daughter nuclei with LETs greater than that
of the incident protons. These reaction products, along with the
ATELLITE borne communication systems have produced
the demand for high-speed, low-power circuits robust
) and drops signifi-
). When a heavy ion
Manuscript received April 4, 2005. This work was supported in part by the
Defense Threat Reduction Agency Radiation Hardened Microelectronics Pro-
gram under MIPR 97-2005 and 00-2062. DFOISR Case Number 04-041, Data
Classification: DTSA Review Exempt by ITAR, 22CFR 125.4(b)(13) Appli-
cable,PaperClearanceNumber04-0153.Thisdocumentcontainstechnicaldata
as defined in the U.S. Government’s International Traffic in Arms Regulations
(ITAR) 22 C.F.R. §120.10 approved for public release and public domain infor-
mation. It is authorized for export under the authority of 22C.F.R. 125.4(b)(13)
andtheDirectorateforFreedomofInformationandSecurityReview(DFOISR).
D. L. Hansen, K. Jobe, P. Chu, and S. F. Meyer are with Boeing Satellite
Systems, Los Angeles, CA 90009 USA (e-mail: david.l.hansen@boeing.com).
P. W. Marshall, M. A. Carts, and C. J. Marshall were with the Naval Research
Laboratory, Washington, DC 20375 USA. They are now with the NASA God-
dard Spaceflight Center, Greenbelt, MD 24528 USA.
R. Lopez-Aguado was with Boeing Satellite Systems, Los Angeles, CA
90009 USA. He is now with Raytheon Company, El Segundo, CA 90254 USA.
Digital Object Identifier 10.1109/TNS.2005.850490
protons themselves, can liberate charge within the transistor’s
sensitive volume causing upsets [5]. The underlying physics by
which protons and heavy ions produce SEU in on-board mi-
croelectronicsis fundamentally similar. Previously,a number of
models have been developed to estimate the proton SEU cross
section from heavy-ion SEU data [5]–[10].
SiGe and InP-based heterojunction bipolar transistors have
gainedprominencefortheirperformanceinhigh-speedapplica-
tions.SiGeprocessingtechnologyhasafairlymatureprocessing
base with high production yields. In addition, the combina-
tion of HBT and CMOS processes to form HBT BiCMOS
technology produces a high level of integration [11]. InP tech-
nology is less mature than SiGe, however, InP based HBTs
havematerialpropertiesthatproveadvantageousforhigh-speed
operation [12].
To date, there have been several experimental [13], [14] and
theoretical [15], [16] studies of the SEU performance of SiGe
circuits.LessworkhasbeendoneforInP[3],[17],[18].Suchin-
formation is essential for space-based applications. We present
here the results of heavy-ion and proton SEU testing of SiGe
and InP shift register circuits. These tests allow us to compare
SEU responses of the devices and to determine the applicability
in the high-speed regime of models predicting proton cross sec-
tions from heavy-ion results for these technologies.
II. EXPERIMENTAL SETUP
Heavy-ion tests were conducted at Brookhaven National
Laboratory (BNL). Devices under test (DUT) were placed in
vacuum (Fig. 1) with RF and dc connections accomplished by
means of feedthroughs containing SMA and 40 pin IDC con-
nections.Thedatapatternwas a
number (PRN) code having every possible 7-bit pattern except
seven consecutive zeros with a nearly even mixture of zeroes
and ones. An HP 71612A pulse pattern generator/bit error
rate test (PPG/BERT) pair was used to supply clock and data
to a differential panel that supplies the DUT with data and
data-not as well as clock and clock-not signal. DUT output is
fed into the error detector on the BERT and a PC is interfaced
with the BERT in order to record the total number of upset
events, the number of bits in error for each event, and the
device current during the test. Care was taken to ensure that
clock and data inputs had optimal time alignment. A series of
different ion beams ( Li,
C,
and device angles were used to produce LET values between
0.3 and 114 MeV cm mg
.
Proton testing was conducted at Crocker Nuclear Labora-
tory (CNL), University of California, Davis, using their 76”
bitslongpseudo-random
Si, Cl,Ni,Br, and I)
0018-9499/$20.00 © 2005 IEEE

Page 2

HANSEN et al.: STUDY OF THE SEU PERFORMANCE OF InP AND SiGe SHIFT REGISTERS1141
Fig. 1. Experimental setup.
isochronous cyclotron. The data was taken with monoenergetic
beams of 25, 38, and 63 MeV. Each 16-bit shift register was
testedinthesametestfixture,atnormalincidentangleandroom
temperature. Setup for the proton test was similar to that for the
heavy-ion tests with a few notable exceptions. The pattern gen-
erator of an Anritsu MP1758A/MP1764A BERT pair was used
instead of the HP BERT; and because protons used produce sig-
nificant secondary radiation, experimenters are required to be
approximately 20 m away from the end of the beam line. Thus
the equipment at CNL was operated remotely.
It should be noted that the LET values listed for the heavy
ion tests in MeV cm mg
are calculated for ions in Si. This
proves to be a convenient unit of measure since it is in common
use, and allows us to compare the effects of identical ions on
different circuits. While the LET in Si is nearly identical to that
in SiGe, the LET in InP is different, particularly at higher LET
values (Table I). However, the differences do not change the
results substantially. At each data point, a sufficient number of
upsetswererecordedsothatthestatisticalcountingerrorateach
point is small. For the heavy-ion data, error bars are estimated
by calculating the standard deviation of multiple runs at a single
LET.Fortheprotondata,theerrorbarsrepresentthesquareroot
of the number of counts.
The test circuits were fabricated using the IBM 5HP SiGe
BiCMOS [19], [20] and HRL Laboratories’ InP G1 [12], [21]
technologies. Parts were mounted onto test carriers and con-
nected to a substrate to facilitate RF and dc connections. To en-
sure adequate penetration by the incident ions, devices had only
a minimal amount of passivation. Two samples from SiGe and
InP technology were tested. The SiGe parts operated at clock
speed ranges of 0.1–6.4 GHz, while the InP were clocked at
0.1–7 GHz. Both circuits use
A block diagram for the two circuits is shown in Fig. 2.
V and 130 mA dc supplies.
TABLE I
LET IN SIGE AND INP FOR DIFFERENT IONS
TheSiGecircuitsarefabricatedfromthreedifferentsizetran-
sistors with 1.0
0.5m , 2.5
emitters. The InP circuits are fabricated using transistors having
2
2 m and 25 m emitters. The SiGe and InP shift reg-
isters are nearly identical, however the InP transistors have a
larger lateral area, leading to larger areas in the layout and con-
sequently, longer line lengths. To compensate for this, the InP
circuits have additional buffers in the data and clock input as
wellasinthelatches.Asaresult,excludingthetransistorsinthe
bias circuitry, the SiGe shift register has 399 transistors, while
the InP shift registers have 560 transistors. If we sum the area of
the base collector junctions for all transistors in each circuit, we
find that the total area is more than 13 times larger for the InP
circuit than for the SiGe circuit. Bias circuitry transistors were
not included in the transistor count because circuit simulations
and subsequent microbeam testing [24] of these devices indi-
cated that the bias circuitry did not upset and therefore does not
0.5m and 5.0 0.5m

Page 3

1142 IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 52, NO. 4, AUGUST 2005
Fig. 2.Block diagram of16-bit shift register.Three of the outputs monitored intermediate stages of the shift register and were notused for data taken in this paper.
contributetothetotalSEUcrosssection.Note,however,thatmi-
crobeamtestLETswerelimitedto8MeV cm mg
tobeam limitations.Whileitis possiblethatthebias andcurrent
sourcecircuitrytransistorswillupsetatthehigherLETsusedfor
these tests, it is reasonable to assume that at higher LETs, they
will continue to be significantly less prone to upset than other
parts of the circuit.
orlessdue
III. RESULTS AND DISCUSSION
A. Heavy-Ion Data
Fig. 3 shows heavy ion data plotted with event cross section
and LET on the vertical and horizontal axes. The event cross
section measures the number of ion-circuit interactions, and is
independent of the number of data bits upset by an ion. We
are distinguishing here the difference between disturbing data
(information) bits and physical memory bits associated with a
physical location. Multiple-bit upsets as used traditionally in
the radiation community, for example, refer to a single particle
strike that hits several physical sites due to the particle’s trajec-
tory across a circuit. In high-speed circuits, the charge injected
at a single location affects multiple clock cycles and therefore
impacts multiple data bits, temporally. To differentiate between
the two mechanisms and minimize confusion about multiple bit
upsets, we have chosen to refer to the disturbance of a train of
data bits as an event, and to measure event cross sections to un-
derstand physical vulnerability of the circuit. Counting upsets
in this manner, the increased vulnerability at higher frequency
arises from the fact that the circuit spends more time in a vul-
nerable state when there are more clock edges [4].
These circuits show fairly typical cross section curves
(Fig. 3) in that the cross section increases rapidly at low LET
values and approaches a high-LET saturation cross section
(
). Discontinuities are present between the data points
collected with different ions. The discontinuities are within the
error bars, thus no corrections were applied to the data. In fact,
correctionmethodssuggestedin[22]increasethevalue ofevent
Fig. 3.
heavy ion irradiation. Symbols indicate the ion type used at each data point. InP
and SiGe data were collected at 7.5 and 4 GHz respectively. Note the scales are
linear. Data points marked by ? represent data for InP at 4 GHz.
Event cross section versus LET for the InP and SiGe circuits under
cross-section data for measurements made at nonnormal inci-
dence angles, making discontinuities larger. Our results suggest
that the cross section for SiGe 5HP parts does not follow tradi-
tional rectangular parallel-piped scaling. This agrees with the
conclusions of previous tests [13]. The threshold LET (
was defined as the LET value for at
proves to be useful since
proton and heavy-ion cross sections, as will be discussed later.
For the SiGe circuit operating at 4 GHz, we measure a
1.78 MeV cm mg
, while for the InP operating at 7.5 GHz
circuit we measure
We determinedby taking an average of the cross sec-
tions measured at the highest LET data points. For InP at
7.5 GHz and SiGe at 4 GHz,
and 8.9
cm , respectively. In order to make a more
direct comparison, data points taken using the InP device at
4 GHz with a LET of 24.7 and 39 MeV cm mg
),
. This technique
is used in models comparing
of
MeV cm mg
are about 1.5cm ,
are plotted

Page 4

HANSEN et al.: STUDY OF THE SEU PERFORMANCE OF InP AND SiGe SHIFT REGISTERS1143
Fig. 4.
ion irradiation. Symbols indicate LET at each data point. Note the scales are
linear.
Event cross section versus data rate for the SiGe circuit under heavy
Fig. 5.
irradiation. Symbols indicate LET at each data point. Note the scales are linear.
Event cross section versus data rate for the InP circuit under heavy ion
(Fig. 3) and suggest that at identical frequencies, the InP and
SiGe shift registers have similar cross sections. This result is
noteworthy since no attempt was made to design SEU hardness
into these parts. In addition, because the InP shift register has
more transistors and thus a larger lateral area than the SiGe shift
register while operating with the same voltage and current the
individual transistors within the InP shift register have, on av-
erage, less current flowing through them. The difference in size
and current should cause the InP cross section to be greater than
that for SiGe [23]. The fact that the InP shift-register cross sec-
tion is similar to that for the SiGe shift register suggests that
relative to InP, SiGe transistors may be slightly more sensitive
to SEU.
To investigate further, in Fig. 4, the SiGe cross section is
plotted as a function of the data rate, the same plot is shown for
InP in Fig. 5. As expected, cross section trends higher with data
rate. In addition, the cross section does not trend toward zero
as the data rate approaches zero. This is similar to previous re-
sults in GaAs HIGFET devices [4] where similar behavior was
attributed to a combination of regions with cross sections that
were data-rate dependent and data rate independent. In Fig. 5,
The InP cross sections are nearly identical at LETs of 24.7 and
39MeV cm mg
,thisresultsfrom thefactthatattheseLETs,
Fig. 6.
function of LET for the InP and SiGe circuits under heavy ion irradiation. InP
and SiGe data were collected at 7.5 and 4 GHz, respectively. Inset graph gives
greater detail of the low LET regime.
Average number of bits in error for each ion-interaction event as a
Fig. 7.
function of data rate for the SiGe circuit under heavy ion irradiation.
Average number of bits in error for each ion-interaction event as a
the cross section is near saturation, and is changing less rapidly
than at LET values near the threshold.
Sincesystemsengineers needtoknownotonlytheupset rate,
but also how many data bits are affected in one upset, we were
careful to look at both the number of events that occurred, and
the number of bits affected by ion strikes. This is extremely im-
portant for mitigation purposes to determine error correction al-
gorithmlengthsrequiredtoachievepropersystemperformance.
Figs. 6–8 show the average number of bits affected by each
heavy-ion interaction event. In Fig. 6, the SiGe circuit shows
an increasing number of bit errors per event with increasing
LET. This is an expected consequence of the fact that as LET is
increased, the amount of charge deposited increases and more
time is required to dissipate the charge.
InPshowssimilarbehaviorexceptatlowLETvalues.AtLET
near
where the cross section is orders of magnitude lower
than saturation, the number of bit errors per event is highest.
At an LET of 1 MeV cm mg
rupt only a single bit, and about 29% corrupt four or more bits.
In contrast, at an LET of 6.8 MeV
the upsets corrupt a single bit, and about 10% corrupt four or
more bits. Thus, at low LET, the average number of bits upset
during an event decreases (Fig. 6), while the event cross sec-
tion increases (Fig. 3). The number of bits in error can be a
about 43% of the upsets cor-
cm mgabout 85% of

Page 5

1144 IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 52, NO. 4, AUGUST 2005
Fig. 8.
function of data rate for the InP circuit under heavy ion irradiation. Scale is
identical to that in Fig. 7 to facilitate comparison.
Average number of bits in error for each ion-interaction event as a
reflection of either the duration of the event or the location of
the upset. In the former case, multiple data bit upsets reflect the
fact that charge introduced to the circuit by an incident ion dis-
sipates according to the RC constant of the circuit node. This
disrupts the data for an amount of time equivalent to the tran-
sient width which can be several clock (or data) bit widths. In
contrast,upsetsoftransistorsintheclocktreecanaffectmultiple
flip-flopsduringasingleclockperiod(Fig.2).Singleeventtran-
sients (SET) in the clock tree are fanned out through the clock
tree buffers creating metastabilities in multiple flip-flops, and
corrupting their data contents. While the long transients require
particles with higher LET, any particle upsetting a clock buffer
can corrupt multiple bits. In Fig. 6, at the low end of the LET
spectrum, charge deposited, and consequently upset time, is a
minimum. Thus this behavior is best explained by the fact that
at low LET, only the transistors in the clock tree are vulnerable
to upset. Categorizing the clock tree circuitry as most sensitive
transistors in the circuit is consistent with the fact that they are
designed to drive minimum capacitance with maximum speed.
Subsequent microbeam testing showed similar results [24]. As
LETincreases,transistorsintheclocktreearenolongertheonly
ones vulnerable. Transistors in the flip-flops, data-input buffers,
and data-output buffers effect only a single bit location in the
data stream during a given clock period, and comprise the ma-
jority of the transistors in the circuit. Although this explains the
behavior of the InP shift registers, it does not explain why the
behavior is not also seen in the SiGe shift register. We note that
the InP shift register has an additional emitter-follower in the
clock tree. Thus the simplest explanation for the difference in
performance is the difference in design.
Figs. 7 and 8 show the average number of bit errors per event
as a function of frequencyata constant LET.InP data is approx-
imately constant up to a frequency of about 3 GHz after which
it increases to a value of almost 2 at 8 GHz. For the SiGe parts
below about 1 GHz, the average number of bits upset during an
interactionisapproximatelyconstantandconvergestosinglebit
errors. The increase above 1 GHz is more rapid at higher LET.
From Figs. 7 and 8, we see that heavy ion upsets in SiGe cir-
cuits typically last longer than those in the InP circuits. The InP
circuits have a higher bandwidth than the SiGe circuits. This
Fig. 9.
proton irradiation. Proton energy is indicated by the symbol. Inset graph gives
greater detail of the InP cross section. Note the scales are linear.
Event cross section versus data rate for InP and SiGe circuits under
allows the charge deposited by an ion strike to be swept away
morequickly,thus,onaverage,theSETcorruptsfewerdatabits.
B. Proton Data
The shift registers were also tested for SEU during proton
irradiation. Both the SiGe and InP devices were irradiated to a
total dose in excess of 3 Mrad without significant degradation
in performance. This is in good agreement with the published
results for total-dose testing [25]–[27]. Event cross section as
a function of data rate during proton irradiation is shown in
Fig. 9. A number of features are readily apparent. The SiGe
cross section shows a pronounced jump near the bandwidth
limit of the device. This is similar to the results seen previously
[2]. The InP shift register was not tested near its bandwidth
limit and cross section remains nearly constant over the entire
frequency range. In the case of clocked devices, such as shift
registers, while the clock tree is always vulnerable, the data
stream—which accounts for most of the transistors in the
circuit—is only vulnerable during the set-up and hold time
of the flip-flops. The higher bandwidth of the InP circuit not
only gives shorter duration upsets, but also results in shorter
set-up and hold times. Thus for the InP circuit, the temporal
window of vulnerability is smaller at higher frequencies. At
all frequencies, the InP circuits have a lower interaction cross
section than SiGe circuits. The SiGe shift registers have a
cross section more than three times higher, at 1 GHz, and
more than eight times higher at 3 GHz.
Multiple data-bit upsets were observed in both circuits
(Fig. 10). In previous studies of InGaAs photodiodes, the ab-
sence of multiple bit upsets indicated that direct ionization was
the dominant mechanism [28]. The fact that we see multiple
bit upsets in increments that cannot be explained by upsets to
the clock buffer suggests that the indirect mechanism involving
nuclear fragmentation plays a role in proton-induced SEU.
Because of the different nuclear reaction dynamics in the two
materials, we expect the indirect upset mechanism to play
different roles in the two types of devices. This warrants further
consideration. The InP HBTs are fabricated from a number of
different layers of InP, InGaAs, and AlInAs with a minimum