New lima latw03

Single Event Upset Mitigation Techniques for SRAM-based FPGAs

Fernanda de Lima, Luigi Carro, Ricardo Reis

Universidade Federal do Rio Grande do Sul
PPGC - Instituto de Informática - DELET
Caixa Postal 15064 CEP 91501-970 - Porto Alegre - RS - Brasil
e-mail: <fglima, carro, reis>@inf.ufrgs.br

Abstract and avionics are more and more considering the use of
This paper discusses high level techniques for designing tolerant techniques to assure reliability [2].
fault tolerant systems in SRAM-based FPGAs, without SEU has a peculiar effect in FPGAs when a particle
modification in the FPGA architecture. TMR has been hits the user’ combinational logic. In an ASIC, the effect
s
successfully applied in FPGAs to mitigate transient faults, of a particle hitting either the combinational or the
which are likely to occur in space applications. However, sequential logic is transient; the only variation is the time
TMR comes with high area and power dissipation duration of the fault. A fault in the combinational logic is
penalties. The new technique proposed in this paper was a transient logic pulse in a node that can disappear
specifically developed for FPGAs to cope with transient according to the logic delay and topology. In other words,
faults in the user combinational and sequential logic, while this means that a transient fault in the combinational logic
also reducing pin count, area and power dissipation. The may or may not be latched by a storage cell. Faults in the
methodology was validated by fault injection experiments sequential logic manifest themselves as bit flips, which
in an emulation board. We present some fault coverage will remain in the storage cell until the next load. On the
results and a comparison with the TMR approach. other hand, in a SRAM-based FPGA, both the user’ s
combinational and sequential logic are implemented by
1: Introduction customizable logic cells, in other words, SRAM cells.
Field Programmable Gate Arrays (FPGAs) are When an upset occurs in the combinational logic, hitting
increasingly demanded by spacecraft electronic designers either the logic or the routing, it has a transient effect
because of their high flexibility in achieving multiple followed by a permanent effect, because the SRAM cell
requirements such as high performance, no NRE (Non that implements that logic or controls that routing has
Refundable Engineering) cost and fast turnaround time. In flipped. This means that a transient upset in the
particular, SRAM-based FPGAs are uniquely suited for combinational logic in a FPGA will be latched by a storage
remote missions because of the possibility of being cell, unless some detection technique is used. When an
reprogrammed by the user as many times as necessary in a upset occurs in the user sequential logic, it has a transient
very short period. As a result, SRAM-based FPGAs offer effect, because the fault can be corrected in the next load of
the additional benefits of allowing in-orbit design changes, the cell. Accordingly, the use of SEU mitigation techniques
with the aim of reducing the mission cost by correcting for programmable architecture must take into account these
errors or improving system performance after launch. peculiarities.
SRAM-based FPGA will be the focus of this work, more In order to mitigate SEU in Virtex® family [1] FPGAs,
specifically the Virtex® family [1] from Xilinx. the Triple Modular Redundancy (TMR) with voting
Transient faults, also called Single Event Upset (SEU), technique combined with bitstream scrubbing has been
are the major concern in space applications, with applied [3]. TMR is a suitable technique for SRAM-based
potentially serious consequences for the spacecraft, FPGAs because of its full hardware redundancy property in
including loss of information, functional failure, or loss of the combinational and sequential logic. Previous results
control. SEU occurs when a charged particle hits the from bitstream fault injection [4] and radiation ground
silicon transferring enough energy in order to provoke a bit testing presented in [5] showed that the use of TMR in
flip in a memory cell or a transient logic pulse in the Virtex FPGAs has confirmed the efficacy of the TMR
combinational logic. SEU on devices has become more structure combined with scrubbing, to recover upsets in the
frequent because of smaller transistor features achieved by FPGA architecture. However, the TMR technique presents
the continuous technology evolution. As a result, not only some limitations, such as area overhead, three times more
space applications but also terrestrial applications that are input and output pins and, consequently, a significant
critical such as bank servers, telecommunication servers increase in power dissipation.

In this work we propose an innovative SEU tolerant A full hardware redundancy, the well known TMR
technique for SRAM-based FPGAs to cope with both approach, can also be used to identify the correct value.
problems above described: power and pins overhead Although it presents a larger area overhead compared to
caused by TMR, and the permanent effect of an upset in time redundancy, since it triplicates all the combinational
the user’ combinational logic. This method combines
s and sequential logic, it does not have major performance
duplication with comparison (DWC) and concurrent error penalties, just the voter propagation time, and it does not
detection (CED) based on time redundancy to detect need different clock phases.
permanent faults in the programmable matrix (SEUs in the In the case of SRAM based FPGAs, the problem of
programmable elements). To the author’ knowledge, no
s finding an efficient technique in terms of area,
references were published about high level concurrent performance and power is very challenging, because of the
error detection techniques for FPGAs that take into account high complexity of the architecture. As previously
the permanent effect of a SEU in the user’ combinational
s mentioned, when an upset occurs in the user’ s
logic, except for the TMR approach. combinational logic implemented in a FPGA, it provokes a
very peculiar effect not commonly seen in ASICs. The
2: Previous Work SEU behavior is characterized as a transient effect,
Several SEU mitigation techniques have been proposed followed by a permanent effect. The upset can affect either
in the past years in order to avoid faults in digital circuits, the combinational logic or the routing. The consequences
including those implemented in programmable logic. A of this type of effect, a transient followed by a permanent
SEU immune circuit may be accomplished through a fault, cannot be handled by the standard fault tolerant
variety of mitigation techniques based on redundancy. solutions used in ASICs, such as Error Detection and
Redundancy is provided by extra components (hardware Correction Codes (EDAC), such as Hamming code, or the
redundancy), by extra execution time or by different standard TMR with a single voter, because a fault in the
moment of storage (time redundancy), or by a combination encoder or decoder logic or in the voter would invalidate
of both. An efficient SEU mitigation technique should the technique.
cope with transient faults occurring in the combinational Special techniques should be developed to FPGAs able
logic, and SEUs in the storage cells. In this way, transient to cope with this type of effect. The SEU mitigation
faults in the combinational logic will never be stored in the technique used nowadays to protect designs synthesized in
storage cells, and bit flips in the storage cells will never the Virtex® architecture is mostly based on TMR combined
occur or will be immediately corrected. with scrubbing [3]. The TMR mitigation scheme uses three
Time and hardware redundancy techniques are largely identical logic circuits (redundant block 0, redundant block
used in ASIC [6, 7]. They range from concurrent error 1 and redundant block 2), synthesized in the FPGA,
detection (CED) to correction mechanisms. The use of full performing the same task in tandem, with corresponding
time or full hardware redundancy permits voting the outputs being compared through a majority vote circuit.
correct value in the presence of single upsets. In the case of The TMR technique for Virtex® is presented in details in
time redundancy, the goal is to take advantage of the [3, 4]. Figure 1 shows the TMR scheme. The user’ flip- s
transient pulse characteristic to compare signals at different flop is replaced by a TMR structure composed of three
moments. The output of the combinational logic is latched flip-flops, three voters and multiplexors that are
at three different moments, where the clock edge of the implemented using LUTs, one for each. The combinational
second latch is shifted by the time delay d and the clock of logic and input and output pins are also triplicated to avoid
the third latch is shifted by the time delay 2d. A voter any single point of failure inside the FPGA, as illustrated
chooses the correct value. The area overhead comes from in figure 1. In this way, any upset inside the matrix can be
the extra sample latches and the performance penalty is voted out by the TMR structure assuring correct values in
measured by twice the duration of the transient upset pulse. the output.
Sequential logic pad
Combinational logic
pad voter tr0
clk0
tr0 tr0 check0

pad

voter tr1
pad clk1
tr1 tr1 check1

pad

pad tr2
clk2 voter
tr2 tr2 check2

Figure 1 - Triple Modular Redundancy for Xilinx FPGAs (User’ combinational and sequential logic and output pads)
s

TMR is an attractive solution for FPGAs because it be used to detect transient upsets, and not transient upsets
provides a full hardware redundancy, including the user’ s that become permanent, as in the case of SRAM based
combinational and sequential logic, the routing, and the I/O FPGAs. Second, in the FPGA, it is not only sufficient to
pads. However, it comes with design penalties, because of detect an upset, but one also must be able to vote the
its full hardware redundancy, such as area, I/O pad correct value in order to assure the correct output without
limitations and power dissipation. Although these interruption in the operation.
overheads could be reduced by using some SEU mitigation The insertion of a CED block in each redundant
solutions such as hardened memory cells, EDACs module of the duplication approach can help to identify
techniques and standard TMR with single voter, these which module is faulty. Figure 2 presents the DWC
solutions are costly because they require modifications in approach able to perform concurrent error detection
the matrix architecture (NRE cost). (CED), called hot backup DWC-CED. The important
Techniques to detect permanent faults in FPGAs have characteristic of this approach is that there is always a
been discussed previously in [8]. However, the authors do correct value in the output of the scheme in presence of a
not discuss about the permanent effect of a single upset in single fault, because the mechanism is able to detect the
the user’ combinational logic, but physical faults that can
s faulty module and to select the correct output out of two.
not be corrected by reconfiguration. When a physical fault Combinational logic
occurs in the matrix, it is necessary to reconfigure the Sequential logic
design avoiding the faulty CLBs in order to continuing dr0
using the part. In this paper, we present a new high level TMR
CED
technique that combines the time and hardware redundancy
with some extra features able to cope with the upset effects
in FPGAs to reduce the number of I/O pads and area CED
overhead for the user’ combinational logic. This method
s Permanent
dr1
copes with the permanent effect of a SEU in the fault detection
programmable matrix. In addition, this proposed method is
also able to detect physical faults, which are permanent Figure 2 – DWC-CED for combinational logic
faults that are not corrected by reconfiguration.
There are many methods to implement a CED,
3: REDUCING PIN AND AREA OVERHEARD BY basically based on time or hardware redundancy. It is
USING TIME AND HARDWARE REDUNDANCY apparent that hardware redundancy is not the appropriate
choice for CED, because it would increase the area
There are always some kinds of penalties to be paid overhead, that in the end it would be larger than the
when designing fault tolerant systems. The penalties come original TMR. Consequently, time redundancy is the only
from the redundancy. Full hardware redundancy approach way to detect a fault without area overhead. The basic
can produce large area, pin and power overheads. Time concept of time redundancy is the repetition of
redundancy can provoke interruption of the system in computation in a way that allows the errors to be detected.
presence of faults and performance penalties. In order to The problem of detecting faults in SRAM-based FPGAs is
explore the design space, one could wonder what would that common time redundancy methods able to detect
happen if some hardware redundancy blocks are replaced transient faults are not applicable in this case, because an
by time redundancy. upset in the programmable matrix has a permanent effect,
Aiming to reduce the number of pins overhead of a full changing the construction of the user’ combinational
s
hardware redundancy implementation (TMR), and at the logic. Consequently, after the occurrence of an upset in the
same time coping with permanent upset effects, we present programmable matrix, the upset will only disappear after a
a new technique based on time and hardware redundancy reconfiguration that can be complete or partial. The upset
to protect the user’ combinational logic, where the double
s can occur at any time, even the propagation time, and it
modular redundancy with comparison (DWC) is combined must be detected.
to a time redundancy upset detection machine, able to To allow redundancy to detect permanent faults, the
detect upsets and to identify the correct value to allow repeated computations are performed differently. During
continuous operation. The sequential logic continues to be the first computation at time t0, the operands are used
protected by TMR to avoid accumulation of faults as directly in the combinational block and the result is stored
previously described, since the scrubbing does not change for further comparison. During the second computation at
the content of a user’ memory cell.
s time t0+d, the operands are modified, prior to use, in such
The possibility of applying concurrent error detection a way that errors resulting from permanent faults in the
based on time redundancy combined with hardware combinational logic are different in the first calculation
redundancy technique for FPGAs looks interesting for than in the second and can be detected when results are
upset detection and upset voting, but there are two
problems to be solved. First, the DWC technique can only

compared. These modifications are seen as encode and faults or with a single fault. The method just needs one
decode processes. The scheme is presented in figure 3. clock cycle in hold operation to detect the faulty module,
Many techniques to encode and decode were proposed and after that it will operate normally again without
in the literature to detect permanent faults [9], such as bit- performance penalties. The final clock period is the
wise inversion for stuck at fault problems, parity prediction original clock period plus the propagation delay of the
for logic functions, re-computing with shift operands output comparator, as presented in figure 6. Sample
(RESO) for faulty arithmetic slices and re-computing with registers (dr0 and dr1) are latched at rise clock edge and
swapped operands (REWSO) for faulty chips. This paper the user’ TMR registers are latched at rise clock+d edge.
s
will show the first case study, a multiplier. For this circuit, Figure 4 shows the scheme proposed for an arithmetic
RESO technique can be successfully used, as presented in module, in the present case study: a multiplier. There are
[10]. two multiplier modules: mult_dr0 and mult_dr1. There are
Time t=t0 multiplexors at the output able to provide normal or shifted
operands. The computed output computed from the normal
Combinational
logic operands is always stored in a sample register, one for each
module. Each output goes directly to the input of the user’ s
comparator

error
TMR register. Module dr0 connects to register tr0 and
clk module dr1 connects to register tr1. Register tr2 will
Time t=t0+d
receive the module that does not have any fault. For
Combinational decoder
default, the circuit starts passing the module dr0. A
encoder
logic comparator at the output of register dr0 and dr1 indicates
an output mismatch (Hc). If Hc=0, no error is found and
output the circuit will continue to operate normally. If Hc=1, an
Figure 3 - Time redundancy for permanent fault detection error is characterized and the operands need to be re-
computed using the RESO method to detect which module
The concurrent error detection method RESO uses the has the permanent fault. The detection takes one clock
principle of time redundancy where the coding function is cycle. While the circuit performs the detection, the user’ s
the left shift operation and the decoding function is the TMR register holds its previous value. When the free
right shift operation. Thus, in the first computation, the faulty module is found, register tr2 receives the output of
operands are computed and stored in a register. At the this module and it will continue to receive this output until
second computation, the operands are shifted k bits to the the next chip reconfiguration (fault correction).
left, computed and the result is shifted k bits to the right DB DA DB DA
(2k bits, if a multiplier or divider). In the proposed DB shifted DA shifted DB shifted DA shifted
application, the operands are shifted by 1 bit. The result of
the second step is compared to the previous result stored in ST0 1 0 0 1 ST0 ST1 1 0 0 1 ST1
the register. A mismatch indicates the presence of a
permanent fault in the module.
Mult_dr0 Mult_dr1
The combination of DWC technique and CED blocks
enables one to detect permanent faults provide a new high- dr0 shifted dr1 shifted
level SEU mitigation technique for FPGAs. In the scheme
presented in figure 4, two clock cycles are necessary to
identify a permanent fault in the combinational logic clk dr0 clk dr1
module. However, this extra time does not occur at every
= = =
clock operation in our approach. Using DWC combined
Hc Tc0 Tc1
with CED for permanent faults, it is possible to take Tc0 Hc Tc1
dr0 dr1
advantage of the simple comparison at the output of the
permanent
duplication scheme to inform whether it is necessary to re- fault detection
compute the data for permanent fault detection. The re-
computation is needed only when a mismatch of the ST
outputs occurs. tr0 tr1 tr2 clk+d

If an output mismatch occurs, the output register will
hold its original value for one extra clock cycle, while the MAJ MAJ MAJ
CED block detects the permanent fault. After this, the
output will receive the data from the fault free module until
the next reconfiguration (fault correction). The important
Figure 4 - DWC-CED proposed scheme
characteristic of this method is that it does not incur
performance penalty when the system is operating free of

Note that in both methods, TMR and the proposed approach and TMR were also performed using Xilinx
technique, the upsets in the user’ combinational logic are
s implementation tools. We use multipliers as case studies.
corrected by scrubbing, while upsets in the user’ s Fault injection in VHDL was used to characterize and
sequential logic are corrected by the TMR scheme used in validate the technique. The fault injection system is able to
the CLB flip-flops. It is important to notice that for upset randomly choose the instant of insertion of the fault, the
correction the scrubbing is performed continuously, to faulty node and the redundant module (mult_dr0 or
assure that only one upset has occurred between two mult_dr1). There is a reset fault signal that works as a
reconfigurations in the design. Some constraints must be scrubbing cleaning up the fault. The circuit chosen for this
observed for the perfect functioning of the technique, same first evaluation was an 8x8 bits multiplier with a register in
as TMR: there must not be upsets in more than one the output. In this way it was possible to inject a set of
redundant module, including the state machine detection faults in each user’ combinational nodes of the example
s
and voting circuit, consequently it is important to use some circuit (193 nodes in total) covering all time span of the
assigned area constraints to reduce the probability of short clock cycle and to emulate the scrubbing between faults.
circuits between redundant module dr0 and dr1. The The multiplier input vectors were also randomly generated.
scrubbing rate should be fast enough to avoid accumulation The faults are injected in a SRAM cell located in the fault
of upsets in two different redundant blocks. Upsets in the node and its effect is permanent until the chip
detection and voting circuit do not interfere in the correct reconfiguration (scrubbing). Fault injection results show
execution of the system, because the logic is already the reliability of the presented method. There were inserted
triplicated. In addition, upsets in the latches of this logic 160 faults in a random node (ranging from 0 to 192) at a
are not critical, as they are refreshed in each clock cycle. random instant at the clock cycle in an 8x8 bits multiplier.
Assuming a single upset per chip between scrubbing, if an All faults were detected and corrected on the fly within one
upset alters the correct voting, it does not matter, as far as clock cycle delay before being captured by the register.
there is no upset in both redundant blocks. Figure 5 shows a graphics representing a fault
In the proposed method, the area reduced by the design occurrence in module dr0 (st_perm_dr1=0, indicating that
compared to the TMR is the area of one user’ s dr0 is fault free, number 1 in fig.), consequently, trv2
combinational logic module and the number of inputs that receives dr1 (mux_select=1, number 2 in fig.). Note that
is reduced from 3 times to 2 times the original number. the value stored in the user’ TMR register is held for one
s
This technique can be used as an additional option for the cycle (number 3 in fig.), while the scheme identifies the
TMR technique for designing reliable circuits in FPGAs free faulty module.
with pads and power reduction. Because the combinational Table 2 presents area results of 8x8 and 16x16 bits
circuit is just duplicated, inputs and outputs can be multipliers, implemented in the XCV300 FPGA using no
duplicated instead of triplicated, as in the TMR approach. fault tolerance technique, TMR technique and the proposed
However, it is important to notice that the TMR inputs technique (DWC-CED for permanent faults). All of the
related to the user’ sequential logic used in the CLB flip-
s multipliers are synthesized with a register at the output.
flops are not changed as triple clocks, reset, etc. Results show that according to the size of the
In addition, the advantage of using this technique is not combinational logic block, it is possible to not only to
only focused in reducing the pin count and the number of reduce the number of I/O pins but also area. Note that the
CLBs, but also in other type of radiation effects such as 16x16 bits multiplier protected by TMR could not be
total ionization dose as this method has the important synthesized in the prototype board that uses a Virtex part
characteristic of detecting permanent faults. So far, we with 240 I/O pins (166 available for the user); while the
have mentioned only SEUs that happen in the SRAM same multiplier, implemented by the proposed technique
programmable cells that are permanent until the next could fit in the chip, and also occupying less area.
reconfiguration. However, a circuit operating in the space According to the user’ application requirements, the
s
environment can suffer from total ionization dose and other designer will be able to choose between a full hardware
effects that can provoke permanent physical damages in redundancy implementation (TMR) or a mixed solution,
the circuit. where time redundancy is combined with hardware
redundancy to reduce pins and power dissipation in the
4: RESULTS interface. It is possible to use DMR and time redundancy
The proposed DWC-CED technique for permanent only in the interface of the FPGA, in this way reducing
fault detection was first validated by fault injection pins. DMR and time redundancy can also be used in the
methodology in a prototype board using emulation. The design to reduce not only number of I/O pads, but also area
fault injection system described in VHDL was specifically for large combinational circuits as presented in
developed to test the proposed technique. Results were table 2.
emulated in an AFX-PQ249-110 board using a XCV300
part. Some area comparisons between the proposed

1

2
3

Figure 5 - Upsets emulation in the Chipscope Analyzer from Xilinx using the Virtex FPGA prototype board
Table 2 - Multiplier Implementations XCV300-PQ240 (* I/O pins were out of range, the part XCV300-BG432 was used.)
No protection TMR DWC-CED
Multipliers 8x8 16x16 8x8 16x16 8x8 16x16
Combinational Input Pins 16 32 48 96 32 64
Sequential Input Pins 2 2 12 12 12 12
Output Pins 16 32 48 96 48 96
Total I/Os 34 66 108 208* 92 172
Number of 4-input LUTs 176 736 576 2,304 532 1,829

5: CONCLUSIONS 7: REFERENCES
This work presented a new technique for upset
[1] XILINX Inc. Virtex™ 2.5 V Field Programmable Gate
detection and voting that combines duplication with Arrays, Xilinx Datasheet DS003, v2.4, Oct. 2000.
comparison (DWC) with concurrent error detection (CED) [2] NORMAND, E.,“Single Event Upset at Ground Level”, IEEE
based on time redundancy for the user’ combinational
s Transactions on Nuclear Science, VOL. 43, N. 6, Dec. 1996.
logic in SRAM-based FPGAs. This technique reduces the [3] CARMICHAEL, C., “Triple Module Redundancy Design
number of input and output pins of the user’ s Techniques for Virtex Series FPGA”, Xilinx Application
combinational logic. In addition, it can also reduce area Notes 197, v1.0, Mar. 2001.
when large combinational blocks are used. The proposed [4] LIMA, F.; CARMICHAEL, C.; FABULA, J.; PADOVANI,
approach was validated by fault injection experiment in a R.; REIS, R., “A Fault Injection Analysis of Virtex® FPGA
TMR Design Methodology”, Proc. of Radiation and its
Virtex prototype board using emulation. A large number of
Effects on Components and Systems (RADECS), Sept. 2001.
upsets were randomly inserted in the user’ combinational
s [5] CARMICHAEL, C.; FULLER, E.; FABULA, J.; LIMA, F.,
logic nodes to emulate faults in the logic. The fault “Proton Testing of SEU Mitigation Methods for the Virtex
injection procedure was developed in VHDL and it FPGA”, Proc. of Military and Aerospace Applications of
represents the effect of a SEU in a SRAM-based FPGA, Programmable Logic Devices MAPLD, 2001.
where it has a transient effect followed by a permanent [6] DUPONT, E.; NICOLAIDIS, M.; ROHR, P., “Embedded
effect. Experiments in 8x8 bits multiplier has shown that Robustness IPs for Transient-Error-Free ICs”, IEEE Design
100% of the faults can be detected all of them can be voted and Test of Computers, May-June, 2002, pp. 56-70.
on the fly before being captured by a CLB flip-flop. [7] ANGHEL, L.; NICOLAIDIS, M., “Cost Reduction and
Evaluation of a Temporary Faults Detecting Technique,”
Ongoing work includes performing additional fault
Proc. 2000 Design Automation and Test in Europe
injection evaluation in more complex circuits such as Conference, ACM Press, New York, 2000, pp. 591-598.
larger multipliers, filters and to study the applicability of [8] YU, Shu-Yi; MCCLUSKEY, E., “Permanent Fault Repair for
this method for other cores such as microprocessors. FPGAs with Limited Redundant Area”, IEEE International
Symposium on Defect and Fault Tolerance in VLSI Systems,
6: ACKNOWLEDGMENTS 2001.
The authors wish to thank Xilinx and CNPq [9] JOHNSON, B. W.; AYLOR, J. H.; HANA, H., “Efficient Use
Brazilian Agency for their support of this work and Renato of Time and Hardware Redundancy for Concurrent Error
Hentschke from UFRGS for his contribution on the case Detection in a 32-bit VLSI Adder,” IEEE Journal of Solid-
State-Circuits, pp. 208-215, Feb. 1988.
study circuit description. [10] PATEL, J.; FUNG, L., “Multiplier and Divider Arrays with
Concurrent Error Detection”, Proceedings of FTCS-25, Vol.
3, 1996.

New lima latw03

More Related Content

What's hot (8)

Viewers also liked (14)

Similar to New lima latw03 (20)

Recently uploaded (20)

New lima latw03