Sie sind auf Seite 1von 4

ASC ’98, PREPRINT, September 13, 1998 1

Timing Jitter and Bit Errors in a 64-Bit Circular


Shift Register
Andrea M. Herr
Department of Physics and Astronomy, University of Rochester, Rochester, New York 14627

Marc J. Feldman and Mark F. Bocko


Department of Electrical and Computer Engineering, University of Rochester, Rochester, New York 14627

Abstract — The bit-error rate of a 64-bit single-flux-quantum circuit operation because once the SFQ pulses are introduced at
circular shift register, operating at a clock frequency of 10-16 GHz the circuit inputs they travel through the circuit at high speed.
was measured. Error incidence depends on the values of the clock The clock period does not affect the occurrence of hold time
and data bias currents and on the clock frequency. Timing viola- errors since it is only the relative position of the data to the pre-
tion arising from thermal jitter is the dominant error mechanism. ceding clock pulse that is critical. Setup time violations, on the
The jitter per JTL stage is estimated to be 340 fs based on the error
other hand, occur when the clock follows the data too closely,
rate data. This corresponds to a noise temperature of 10 K.
i.e. the data comes too late in the clock cycle. As the clock
period decreases, the likelihood of errors due to setup time vio-

D
I. I NTRODUCTION lation increases. In effect, any variation in the position of either
IGITAL ELECTRONICS based on the single flux quan- the clock or data pulse, i.e. jitter, increases the hold and setup
tum (SFQ) offers an appealing alternative to conventional time of the SFQ cell.
transistor- based circuits. A feature of this technology is that
II. B IT-E RROR M EASUREMENT
the clock must be delivered to each basic logic element. These
circuits typically have a regular systolic array architecture, in The 64-bit CSR is comprised of eight columns of eight shift
which the data flow through the circuit without feedback (see, register cells, alternately clocked in concurrent and counterflow,
for example, [2]). Circuits with recurrent data paths, in which as shown in Fig. 1; a schematic of the basic shift register cell
the data is required to circulate around closed loops, are subject used in these experiments is shown in Fig. 7 of [8]. The overall
to more stringent timing constraints and must be carefully de- block clocking of the CSR is counterflow; the single exception is
signed with these in mind: the clock skew around a closed loop the long JTL which connects the output of the first column to the
must be zero. Timing errors in recurrent circuits are a likely input of the last (labelled A to D in Fig. 1). There are three bias
cause of circuit failure. leads which power the circuit: one each for the clock and data
Perhaps the simplest recurrent rapid-single-flux-quantum paths, and another for the delay line. Adjusting these individual
(RSFQ) circuit is the circular shift register (CSR). We have dc-current biases changes the relative speed of the clock and
demonstrated a 64-bit RSFQ CSR which showed correct opera-
tion to a clock frequency of 18 GHz, published in [1]. We have BIAS

also measured the bit error rate (BER) of the 64-bit CSR. BER is DATA
defined as the probability that an error in circuit operation will IN CB JTL

occur during a complete circulation of the data through all 64


A B C D
stages of the register, as a function of frequency and bias cur-
rent. We assess the amount of thermally induced timing jitter
present in the circuit by analyzing the errors which are caused
8 STAGES CONCURRENT

8 STAGES COUNTERFLOW

8 STAGES CONCURRENT

8 STAGES COUNTERFLOW

8 STAGES CONCURRENT

8 STAGES COUNTERFLOW

8 STAGES CONCURRENT

8 STAGES COUNTERFLOW

by timing violations. These timing errors are distinct from the


functional errors previously considered in [3], [4], [5]. The shift
register cells used here follow the RSFQ timing convention as
defined in [7]. Therefore, the CSR may be considered a case
study of timing considerations in recurrent SFQ circuits.
Two types of timing errors are possible in our CSR. One oc-
curs when the data pulse follows the clock pulse too closely,
i.e. the data pulse arrives at the input of any one register cell S
too early in the clock cycle, in violation of the hold time of the
S S S S S S S
cell [6]. Hold time violations can occur even during low-speed
DATA OUT

CLOCK IN
Manuscript received September 15, 1998.
This work was supported in part by the University Research Initiative at the
University of Rochester, sponsored by the Army Research Office under Grant Fig. 1. Block diagram of the 64-bit circular shift register. Notation:
No. DAAL03-92-G-0012. CB–confluence buffer, S–splitter.
ASC ’98, PREPRINT, September 13, 1998 2

0
data pulses traveling through the register, making errors more or 10
less likely to occur. The error rate is expected to be a function f = 10.6 GHz
f = 13.4 GHz
of both the clock frequency and these externally supplied dc- f = 15.8 GHz

error rate (per 64 clock cycles)


currents. A more detailed description of the circuit architecture
may be found in [1].
The circuit was fabricated using the standard 1 kA/cm2
−5
10

HYPRES Niobium process [9]. The chip was mounted in a


dip-stick probe and submersed in liquid helium. Two -metal
shields and a Pb shield provided magnetic shielding.
To perform the error rate experiment, the three bias currents −10
10
were first set to values which maximized the circuit margins
for high-speed (15.8 GHz) operation: 30.5 mA for the clock,
25.0 mA for the data, and 3.5 mA for the delay line. These val-
ues define the nominal operating point of the circuit. The cur-
rent supplied to the delay line was held constant throughout the 10
−15

entire error-rate experiment. The clock and data bias currents 29 29.5 30 30.5 31 31.5
clock bias current (mA)
were individually varied to induce errors in circuit operation;
one of these two currents, the test bias, was changed while the Fig. 2. Measured (points) and best-fit error function dependence of
other was held constant at its nominal value. In this way, we BER on clock bias current. The three sets of data correspond to dif-
were able to isolate errors caused by the test current. Systematic ferent clock frequencies. The dashed line represents the nominal bias
adjustment of the supplied bias currents permitted errors to be point.
detected in each of four regimes: high and low clock bias, and
high and low data bias. 10
0

We could not verify the integrity of the data sequence after f = 10.6 GHz
each individual high-speed circulation, since it was not possible f = 13.4 GHz
f = 15.8 GHz
error rate (per 64 clock cycles)

to apply a fixed number of high-speed clock pulses to the regis-


ter input using our on-chip Josephson ring oscillator. Instead, we
−5
applied the multi-GHz clock signal for a relatively long time in- 10
terval before halting the high-speed operation in order to check
the data sequence for errors, as described in [1]. We were able
to determine only whether the data had survived all of the circu-
lations intact.
−10
The circulation time was varied between 190 ms and 10

60000 ms. One hundred experimental trials were performed for


each combination of bias condition and circulation time. The
circulation time was increased from the minimum to the maxi-
mum value, and then subsequently decreased until the minimum −15
value was again reached. Data were taken twice for all but the 10
22 23 24 25 26 27
longest circulation times; the reproducibility of the results was data bias current (mA)
thus confirmed. After each of the 100 individual experiments,
Fig. 3. Measured (points) and best-fit error function dependence of
the integrity of the data output was verified. If the data sequence BER on data bias current. The three sets of data correspond to different
contained errors, the CSR was said to have “failed.” The num- clock frequencies.
ber of observed circuit failures was counted for each set of 100
trials; hence the number of failures ranged from zero to 100. For
those trials which were repeated, the error counts from the two through n circulations is given by
sets of 100 trials were summed.
Automated data acquisition permitted the continuous collec- F = 1 , pn : (1)
tion of experimental data. It took nearly eight hours to obtain the
error counts for each bias point when the circulation time was The quantity F was measured directly in the experiment and is
varied over its full range. Hence, the entire experiment lasted given by
three weeks. The circuit worked continuously over this time pe- F = Nf =T (2)
riod. where Nf is the number of circuit failures observed during T
The BER of the circuit can be derived from the experimental experimental trials for a given combination of test bias and data
data. Let p = 1 , f be the probability that the integrity of the circulation time. It follows that
data is maintained during a single circulation through the entire
64-bit CSR (64 clock cycles). The BER of the circuit is the
1
f = , ln (1 , Nf =T ) (3)
probability, f , that the data sequence does not remain intact. The n
probability that the correct data sequence was not maintained for f  1.
ASC ’98, PREPRINT, September 13, 1998 3

Figs. 2 and 3 show the error rate of the 64-bit CSR as a func- add in quadrature. In the case of the CSR, the resistance R is
tion of the test bias current. The error rate of the CSR is fre- the parallel combination of all of the resistors contributing to
quency independent in the high data bias and high clock bias the noise along the critical path.
regimes. However, there is a fundamental difference between An error in circuit operation will occur if the amount of noise,
the error rate curves in these two regimes. In the case of high In , is large enough to cause the operating margins to be ex-
data bias, the error rate rises as the data bias is increased from ceeded when the circuit is biased with current Ib . That is, an
the nominal operating value. This indicates that the error is ei- error will occur if the magnitude of the noise current exceeds
ther due to a hold time violation or a functional error induced in some threshold value, It = jIm , Ib j, where Im is the value
one of the circuit components. In the high clock bias regime, the of bias current such that in the presence of noise, the error rate
error rate shows no dependence on the magnitude of the supplied of the circuit is 50%; Im is the current margin. It is important
dc-current; the error rate is essentially constant. This indicates to note that the value Im does not correspond to the high-speed
that the error mechanism does not involve timing. We attribute operating margins of the CSR quoted in [1]. In that case, two
these errors to junction over-bias. trials were performed in which the data were made to circulate
The error rate is both frequency and current bias dependent through all 64-stages of the register for 200 ms. This corre-
as the data bias is decreased from its nominal value. This is sponds to more than 30 million circulations at 10 GHz. If the
the unambiguous signature of a setup time violation. Circuit data survived all of these trials intact, then the circuit was said
simulations support this conclusion and indicate the error occurs to have “passed.” The quoted margin is the value of bias current
at point D in Fig. 1. Lowering the data bias current decreases the for which this requirement was fulfilled; this corresponds to an
speed of the data pulses traveling through the circuit, causing error rate of much less than 50%. Assuming that an error oc-
them to come later in the clock cycle than they would at the curred during only one of the 30 million circulations, the error
nominal operating point. The data arrives at the input of the first rate is approximately 3  10,8 .
counterflow register at point D progressively later in the clock If we integrate the Gaussian probability density outward from
cycle until it arrives too close to the following clock pulse and a the threshold current, It , to infinity, we obtain the probability
setup time violation results. This timing violation was observed that the noise current will exceed the threshold value, and con-
to affect the data sequence in the same way each time; the last sequently, the probability that the circuit will fail during any one
“1” in a series of eight “1’s” was delayed by one clock cycle. circulation of the data. Hence, we obtain
Finally, in the low clock bias regime, the error rate shows "r #
1 R
8kB T B jIb , Im j
two types of functional dependence. All three error-rate curves
BER = erfc
are clearly dependent on the magnitude of the current; however, 2 (5)
we notice that the error-rate curves show no frequency depen-
dence until the clock rate reaches some threshold value (between where erfc is the complementary error function.
13 GHz and 15 GHz). This suggests that there are two distinct This theoretical function was fit to the experimental error-rate
error mechanisms involved. At frequencies below the thresh- curves for low clock bias and low data bias currents, for which
old, the error can be attributed to either a hold-time violation or the error mechanism is known to involve timing violations at
a functional error in one of the circuit elements; the signature point D in Fig. 1. Jitter in the position of the clock pulse as it
of hold time violation is that the first “1” in the series of eight arrives at points A and D, and jitter in the path the data travels
“1’s” advanced by one clock cycle. However, as the frequency between points A and D both contribute; it is the relative timing
is increased, the likelihood of a setup time violation somewhere of clock and data which is important. Note that at each counter-
within the circuit increases. flow register stage along the first column of the CSR, the data
III. E FFECTIVE T EMPERATURE is synchronized with the clock, i.e. the data is always released
when the clock arrives at the input of the register cell. Since the
Occasional circuit errors are caused by noise current arising data pulse always reaches the input of the subsequent concurrent
from thermal fluctuations in the shunt and bias resistors or in- cell well in advance of the next clock pulse, small variations in
troduced from external sources. In the following, we will con- the time of arrival of the data pulse at the input of any of the
sider only the effects of Johnson (thermal) noise in the resistors. concurrent registers in the first column do not affect the time at
The amplitude of the noise current is Gaussian distributed with which the data pulse reaches point D. Hence, the jitter which
standard deviation  equal to the root-mean-square noise current causes the timing errors at point D is attributed to current noise
produced by a resistance R at an absolute temperature T : produced by the resistors in the clock path up the first and last
r
=
4kB T B (4)
columns of registers, the clock splitter network between the two
columns, and the data path formed by the long connection from
R A to D. The value of R results from the parallel combination of
where kB is the Boltzmann constant and B is the bandwidth the shunt and bias resistors along this critical path which yields
of the measuring instrument. In our case, the instrument is the a value of 0.0112
, as fabricated.
Josephson junction itself, which is capable of responding to fre- The value of Ib used to fit the curves is the amount of test
quencies from dc to the plasma frequency, fp . For the 1 kA/cm2 current applied to the critical path. In the case of low data bias,
junctions used, fp  137 GHz. The total noise current from this is the amount of bias current supplied to the JTLs on the
multiple independent sources is the square root of the sum of output of the first column and the arm of the confluence buffer
the squares of the individual noise currents; the noise currents traversed by the data pulse, about 9:4% of the total data bias
ASC ’98, PREPRINT, September 13, 1998 4

current. For the case of low clock bias, the relevant bias current to the experimental data for all three error-rate curves. The
is that supplied to the clock lines of the first seven register stages JTL stages in the clock and data paths have Ic in the range
in the first and last register columns (these affect the time at 200-400 A. We estimate the shift register stage to have jitter
which the data is released from the last register in the column) equivalent to that of two JTL stages. The critical path is com-
as well as the current supplied to the relevant arm of the splitter p
prised of the equivalent of 90 JTL stages. Hence, the timing
network that carries the clock pulse to the clock input of the last jitter per stage is approximately t = 90, or 340 fs. Note that
column. This is approximately 27:8% of the total clock bias this is significantly larger than that reported in [10], in which jit-
current. ter was estimated based on linewidth measurements of an RSFQ
The current bias operating margin, Im , and the absolute tem- Josephson ring oscillator.
perature, T , were used as fitting parameters for the theoretical
error-rate function given by (5). An effective temperature of ap- V. C ONCLUSION
proximately 10 K gave the best overall fit to the data in these two Experiments to determine the bit-error rate of a 64-bit CSR as
regimes. This is significantly higher than the physical tempera- a function of operating frequency and clock and data bias cur-
ture of 4.2 K. Additional noise sources may have been present, rents have been performed. The error-rate curves thus obtained
or our simple threshold model may have underestimated the ef- suggest an effective temperature of 10 K. Although this is much
fect of thermal noise in the circuit, which includes nonlinear ele- higher than the bath temperature, it is in good agreement with
ments (the Josephson junction). The 10 K effective temperature previously published results. In addition, we have determined a
is consistent with results presented in [4] and [5]. timing jitter of 340 fs per JTL stage.
We have demonstrated that timing errors due to thermally-
IV. T IMING J ITTER induced jitter can be the dominant error mechanism in a large
SFQ circuit, eclipsing functional circuit errors. While error-rate
Entirely apart from the question of effective temperature, the curves associated with functional errors have a very steep slope
data can be used to determine the amount of timing jitter present as a function of current bias and extrapolate to near negligible
in the 64-bit CSR. To do so, we must quantify how changes in values [4], those associated with jitter-induced timing errors in
the test bias current affect the relative positions of the clock and the CSR have a more gradual slope. On the other hand, tim-
data pulses. ing errors always may be avoided with sufficiently conservative
In the case of low data bias current, we must determine how design; this comes at the expense of operating frequency. It is
the separation time between the conflicting data and clock pulses important to consider the effects of timing jitter when design-
is affected by the data bias current. To this end, numerous cir- ing a large SFQ circuit and to design conservatively in order to
cuit simulations were performed and the separation time mea- attain an acceptable bit-error rate.
sured as a function of data bias current. Using this information,
the error-rate curves shown in Fig. 3, may be transformed so R EFERENCES
that BER is given as a function separation time instead of bias [1] A. M. Herr, C. A. Mancini, N. Vukovic, M. F. Bocko, and M. J. Feld-
current. man, “High-speed operation of a 64-bit circular shift register,” IEEE Trans.
Jitter in the circuit can be estimated again assuming a thresh- Appl. Supercond., vol. 8, pp. 120–124, September 1998.
[2] V. K. Semenov, Y. A. Polyakov, and D. Schneider, “Implementation of
old model; an error will occur if the jitter exceeds the timing Oversampling Analog-to-Digital Converter Based on RSFQ Logic,” Ext.
margins in the circuit. We assume the timing jitter is Gaussian Abstr. 6th Int’l Supercond. Elec. Conf., vol. 2, pp. 41–43, June 1997.
distributed with standard deviation t , the standard deviation of [3] Q. P. Herr and M. J. Feldman, “Error rate of a superconducting circuit,”
Appl. Phys. Lett., vol. 69, pp. 694–695, July 1996.
the quadrature sum of timing jitter in both the clock and data [4] Q. P. Herr and M. J. Feldman, “Error rate of RSFQ circuits: theory,” IEEE
paths. Timing errors will occur if the jitter exceeds a threshold Trans. Appl. Supercond., vol. 7, pp. 2661–2664, June 1997.
tt = jt , tm j where t is the separation time between the con- [5] Q. P. Herr, M. W. Johnson, and M. J. Feldman, “Temperature-dependent
bit-error rate of a clocked superconducting digital circuit” this conference.
flicting clock and data pulses in the absence of noise and tm is [6] K. Gaj, E.G. Friedman, and M.J. Feldman, “Timing of multi-gigahertz
the separation time which results in a 50% failure rate, the “sep- rapid single flux quantum digital circuits,” J. of VLSI Sig. Proc., vol. 16,
pp. 247–276, June–July 1997.
aration time margin.” [7] K. Likharev and V. Semenov, “RSFQ logic/memory family: a new
Following the procedure outlined in Section III we obtain Josephson-junction technology for sub-terahertz-clock-frequency digital
  systems,” IEEE Trans. Appl. Supercond., vol. 1, pp. 3–28, March 1991.
1 1
BER = erfc p jt , tm j :
[8] C.A. Mancini, N. Vukovic, A.M. Herr, K. Gaj, M.F. Bocko, and M.J. Feld-

2 (6) man, “RSFQ circular shift registers,” IEEE Trans. Appl. Supercond., vol.
2t 7, pp. 2832–2835, June 1997.
[9] HYPRES, Inc., HYPRES Niobium process flow and design rules, 175
The standard deviation, t , and the timing margin, tm , were Clearbrook Road, Elmsford, NY 10523, http://www.hypres.com.
[10] V. Kaplunenko, V. Borzenets, N. Dubash, and T. Van Duzer, “Supercon-
used as fitting parameters for the theoretical error-rate function ducting single flux quantum 20 GB/s clock recovery circuit,” Appl. Phys.
given by (6). A standard deviation of 3.2 ps gave the best fit Lett., vol. 71, pp. 128–130, July 1997.

Das könnte Ihnen auch gefallen