Ðåôåðàò: Physical Methods of Speed-Independent Module Design
.
As to choosing
value of rb it must be done with regard to
maximal voltage drop Vmax .
If V>750mV, the diode D1 is in
active mode and while rb <<R1 the
condition Ir <<Id is
true. So, in the large current region IId and
Equation (2) determines an almost linear dependence between I and V. For instance, if the
maximal voltage drop Vmax =900mV
and maximal input current Imax=2mA, then in accordance with
the Equation (2) rb 100. Typical element values for the
OVD circuit with Vth
=400mV are given in Table 1.
The turn-on ton and
turn-off toff delays
of the OVD circuit depend on the OVD itself and the CMOS CL as well. (Switching
the OVD output from low to high voltage is called "turning-on" and
reverse switching is called "turning-off".)
Consider a
piece of CMOS CL and its interaction with OVD circuit (Fig.11). The piece is an
SPP including N logic gates. Each gate is shown symbolically as a
connection of PMOS and NMOS networks. All the capacitances affecting ton and
toff can
be brought down to three components:
(i) CLi is the load capacitance of the
i-th gate;
(ii) Cpsi is
the power supply bus capacitance associated with the i-th gate;
(iii) Cin is
the input capacitance of the OVD circuit.
Let pi is
a probability of the i-th gate being in the state of high output
potential. In this state the capacitance CLi is
connected with power supply bus through the low channel resistance of turned-on
transistors in PMOS network of the i-th gate. Then equivalent
capacitance Ceq connected
to the OVD circuit input equals
(7)
where N is a number of
gates in the considered SPP. Here the resistance of conducting PMOS network is
assumed to be negligible.
Equation (7)
is also true for CL including several SPPs. In that case summing must be
carried out for all the gates belonging to CL.
Simulation
shows that ton and
toff are
proportional to the OVD time constant =R1Ceq. It was also obtained that when N>20,
the component under the sign of summation in Equation (7) can be much larger
than the component Cin. Due to voltage drop V the effective power
supply voltage is reduced and CL performance is decreased by about 35 percent
[7].
In order to
make SIM operating faster special attention must be paid to reducing the
capacitance introduced by CL.
4.3 Speed-independent address
bus
The simplest
case of CL is a scheme degenerated into a set of wires called a multi-bit bus.
Let us develop the OVD circuit for such a CL.
Multi-bit bus
consists of several lines. Each line can be considered as a medium for signal
propagating from one end of the chip to another. Delay of signal propagation
through a line depends on several factors:
(a) output
impedance and symmetry of driver circuit;
(b) initial state
of the line: if driver is symmetrical, line switching from high to low voltage
lasts shorter than reverse switching;
(c) electrical
properties of the line as a signal propagation medium (resistance of conducting
layer and capacitances between the line and other wires next to it);
(d) length of the
line;
(e) input impedance
and sensitivity of receiving circuit.
Since
different lines of the bus operate in different conditions (a)-(e), signal
propagation delays are different, too. From the standpoint of environment the
bus behaves like any other more complicated CL.
Asynchronous
RAM designers use a bus transition detector since 1980s [13-15]. Such a
detector is usually based on double-rail address coding and two series
connected transistors for each address bit [15]. One of the transistors
receives the true address signal and the other receives the complementary
address signal of the particular address bit. For any steady state condition
one of the transistors will be turned on and one will be turned off. There
will be a finite rise and fall time during a transition of the address bit. There
is a short time during which both transistors are conducting. The establishment
of the conductive path provides the detection of the address transition. In
the first asynchronous RAMs the output signal of the transition detector is
used for bit line precharging and for enabling/disabling sense amplifiers
and peripheral circuitry.
Self-timed RAM
announced in 1983 [14] used transition detectors not for address transition
only but also for detecting read/write completion and address/bit line precharge
completion as well.
The CMOS
transition detector was invented in 1986 [15]. This circuit is also based on
double-rail coding and uses a pair of series-connected NMOS transistors
(Fig.12). The scheme for n-bit bus control contains n line
transition detectors (LTDs) and n AND-gates. Outputs of AND-gates are
united in node M forming wired OR. The output inverter serves as a
pulse shaper. Capacitors C1 and C2 are
intended to prolong rise time of the LTD output signal (true and
complementary). This is necessary for reliable detection.
The main
drawback of the circuit is speed dependence. One can see that if true and
complementary address bit signal have different propagation delays, the
conducting path via NMOS transistors will never be formed.
Using the OVD
circuit proposed in Section 4.2 as LTD we can avoid this drawback.
Note that
address transmission through the address bus is unidirectional. So to detect
completion of bus transition it is enough to recognize the bus state at the
destination end. For this purpose we modify CL to consist of n lines.
The modification means introducing n LTDs, each actually a CMOS inverter
chain. Each chain contains two inverters loaded with a capacitance (Fig.13). Input
of each LTD is connected with corresponding line of the bus at the destination
end. Power supply pads of all LTDs are connected to the current input of the
same OVD circuit.
The parameters
of the input current signal for the OVD circuit are varied by
(i) value of
capacitances C1
and C2 ;
(ii) dimensions of
MOS transistors M1 -M4 .
Since all
transitions in CL are of the same duration and can be lengthened to be outlast
the OVD turning-on time, we simplify the interface circuitry by disallowing
the asymmetrical delay.
Due to short duration of
normal transition in this CL we must take into account the integral nature of
the sensitivity of the OVD circuit. OVD sensitivity depends on both amplitude
and width of input current pulse. Simulated operation region of the OVD circuit
for current pulses shorter than 30ns is shown in Fig.14. It is obvious that in
this case the threshold of the OVD circuit must be determined by threshold
charge Qth value.
The OVD input charge Q equals to where
I is OVD input current, t is a moment of time when transition
occurs, w is a width of input current pulse. Turning-on condition for
the OVD circuit is Q=Qth.
When the LTD
circuit shown in Fig.13 is used, the charge value Q is determined by
either C1 or
C2.
Namely, if the line goes from low to high voltage, Q=VC2. If the
line goes in the reverse direction then where
V is charging/discharging voltage,
approximately equal to the effective power supply voltage: VVdd -V. Here Vdd is OVD
power supply voltage and V is CVC voltage drop.
The OVD
circuit with typical parameters (See Table 1) has a threshold charge value Qth =4.010-12 C. When C1 =C2 =CL ,
the minimal value of CL providing OVD capacity for
operation is about 1.010-12 F.
Influence of
transistors M1 -M4
dimensions on LTD delay d is determined by approximation [17]:
where ~ is a sign
of proportionality, Gn and Gp are
the conductances of NMOS and PMOS transistors respectively (CL =C1 =C2.)
Since and where W and L
are width and length of transistor channels of the corresponding conduction
type, the LTD delay d is proportional to .
It has been
obtained that for , , CL=1.0pF
and Vdd-V=5.0V the LTD delay d=7.6ns.
When LTD works
jointly with the OVD in the speed-independent bus, the real value of the LTD
delay will increase by 30-40 percent due to OVD's R1 effect
on the effective power supply voltage.
To determine
the appropriate value of R1 in the OVD circuit we must know
threshold input current Ith corresponding to threshold
voltage drop Vth recommended
to be equal to 400mV.
Average input
current Iav in
transient state of one line is determined by the expression Iav =CLv where v is the
average rate of increase in the output signal for an inverter included in LTD.
For typical values v=1.0109 Volts per second and CL =1.0pF,
Iav =1.0mA.
Accepting Ith =0.4mA
and Imax=2.0mA
we obtain R1=1k and rb=100.
Simulation has
shown that in this case OVD turning-on delay can be approximated by an
empirical expression:
ton[ns]=8.1+0.1n
where n is the address bus
bit capacity. Total delay of recognizing address transition ttot =dg+ton where
g is a coefficient of the LTD delay increase due to reducing power
supply voltage. As we showed above g1.35. It can be seen that if n=32,
ttot=21.6ns.
4.4 Speed-independent adder
The circuit we
use in this Section as a CL was a touch-stone for many speed-independent
circuit designers for about four decades. We mean a ripple carry adder (RCA)
which is actually a chain of one-bit full adders (Fig.14).
Each full
adder calculates two Boolean functions: sum si=aibici and
output carry ci+1=aibi+bici+aici
where ai,
bi
are summands, ci is
input carry and stands for XOR operation.
In 1955
Gilchrist et al. proposed speed-independent RCA with carry completion signal
[18]. In 1960s that circuit was carefully analyzed and improved [19-21]. In
1980 Seitz used RCA for illustrating his concept of equipotential region and
his approach to self-timed system design [4].
Now we use RCA
as a CL for illustrating our approach to SIM design.
As it was
shown in Section 4.2 the turn-on and turn-off delays of the OVD circuit are
proportional to the equivalent capacitance Ceq associated
with OVD circuit input. Capacitance Ceq depends linearly on a number of
gates N in CMOS CL. To speed up a SIM it is necessary to reduce a number
N. This can be reached by structural decomposition CMOS CL into
subcircuits CL1, CL2, etc. Each subcircuit CLi is connected to its own
detecting circuit OVDi or directly to the power supply if this
subcircuit transition does not affect the transition duration in CL as a whole.
Each detecting circuit OVDi generates its own OV signal which is
combined with other OVDs' output signals via a multi-input OR (NOR) element.
The output signal of that element serves as OV signal of the CMOS CL.
Multi-bit RCA
computation time is determined by length of maximal activated carry chain. A
lot of papers were devoted to analysis of carry generation and carry propagation
in RCA [19-21], many of them contained their own methods for estimation or
calculation of average maximal activated carry chain. We do not intend to add
another one.
Let us have a
look inside RCA. As it was mentioned above RCA consists of one-bit full adders
and each full adder consists of two parts: forming sum si part
and forming carry ci+1 part
(Fig.16).
In multi-bit
RCA all forming sum parts do not interact with each other and do not affect on
transition duration in RCA. Each forming carry ci+1 part
receives ci signal
from preceding forming carry part and sends ci+1 signal
to consequent one.
To decompose
RCA we use three heuristic tricks:
(i) All forming sum
parts we connect directly to power supply.
(ii) We divide each
forming carry part into three subcircuits denoted in Fig.16 by numbers 1,2 and
3. All subcircuits 1 we connect directly to power supply because they do not
contain input ci and
so do not contain carry propagation path.
(iii) All
subcircuits 2 we connect to OVD1 and all subcircuits 3 we connect to OVD2.
Outputs of OVD1 and OVD2 are connected to two-input NOR-gate forming RCA OV
signal in positive logic manner (Fig.17).
OVD1 and OVD2
input currents I1 and
I2 curves
for 6-bit RCA and longest transition duration are shown in Fig.18.
Accepting Vth1,2=400mV
we calculated the OVD circuits parameters. It was obtained R11=5k, Ith1=0.08mA,
R12=3k, Ith2=0.13mA.
OVD1 and OVD2 delay dependencies on a number of bits in RCA are shown in
Fig.19.
4.5 Comparison of SIMs with
synchronous counterparts
Transition
duration in CL is a random variable. Probability of transition with duration D
is determined by implemented Boolean function and distribution of input
logical combinations. Domain of possible values for variable D occupies
the interval [0;Dmax].
Here Dmax is
a length of critical path in CL.
Let is a mathematical
expectation of transition duration in CL where Di is
a length of i-th SPP in CL, pi is a probability of i-th
path being the longest activated SPP.
When CL works
in the synchronous mode, the cycle duration Ts is
chosen with regard to maximal transition duration Dmax.
Certain margin must be added to Dmax to provide reliable operation of
CL in the case of CL parameter variations: Ts =kDmax
where k is a margin coefficient.
In SIM cycle
duration is a random variable with expectation Tsi = gDme+toff+tif where
g is a coefficient of CL delay increasing due to reducing power supply
voltage, toff is
turn-off delay of the OVD circuit, tif is an interface circuitry delay.
We determine
efficiency E for speed-independent mode of CL operation as relative
increase of SIM performance in comparison to its synchronous counterpart:.
Generally,
speed-independent mode is more efficient than synchronous one if Ts >Tsi or,
in other words, .
In the case of
RCA where tc is
a delay of carry forming part, n is a number of full adders in RCA.
It has been
shown [19] that in n-bit RCA Dme tclog2(5n/4). Then, in the case
of speed-independent operation Tsi=gtclog2(5n/4)+toff+tif.
We have
obtained dependencies of Ts , Tsi on
a number of bits in RCA that are shown in Fig.20. As it can be seen,
speed-independent operation of RCA is more efficient while n>8.
5.Conclusion
6.Acknowledgement
I would like to
thank Igor Shagurin and Vlad Tsylyov of the Moscow Physical Engineering
Institute for helpful discussions of this work. I am also grateful to Chris
Jesshope of University of Surrey and Mark Josephs of Oxford University who
kindly provided the latest material on their research in the area of
delay-insensitive circuit design.
References
[1] Miller, R.E., Switching
theory (Wiley, New York, 1965), vol.2, Chapter 10.
[2] Unger, S.H., Asynchronous
Sequential Switching Circuits (Wiley, New York, 1969).
[3] Armstrong, D.B., A.D.
Friedman, and P.R. Menon, Design of Asynchronous Circuits Assuming Unbounded
Gate Delays, IEEE Trans.on Computers C-18 (12) (1969) 1110-1120.
[4] Seitz, C.L., System
timing, in: C.A. Mead and L.A. Conway, eds., Introduction to VLSI Systems
(Addison-Wesley, New York, 1980), Chapter 7.
[5] Izosimov, O.A., I.I.
Shagurin, and V.V. Tsylyov, Physical approach to CMOS module self-timing, Electronics
Letters 26 (22) (1990) 1835-1836.
[6] Veendrick, H.J.M.,
Short-circuit dissipation of static CMOS circuit and its impact on the
design of buffer circuits, IEEE J. Solid-State Circuits SC-19
(4) (1984) 468-473.
[7] Chappell, B.A, T.I.
Chappell, S.E. Schuster, H.M. Segmuller, J.W. Allan, R.L. Franch, and P.J. Restle,
Fast CMOS ECL receivers with 100-mV worst-case sensitivity, IEEE J.
Solid-State Circuits SC-23 (1)
(1988) 59-67.
[8] Chu, S.T., J. Dikken,
C.D. Hartgring, F.J. List, J.G. Raemaekers, S.A. Bell, B. Walsh, and R.H.W.
Salters, A 25-ns Low-Power Full-CMOS 1-Mbit (128K8) SRAM, IEEE J.
Solid-State Circuits SC-23 (5) (1988)
1078-1084.
[9] Frank, E.H., and R.F.
Sproull, A Self-Timed Static RAM, in: Proc. Third Caltech VLSI Conference
(Springer-Verlag, Berlin, 1983) pp.275-285.
[10] Donoghue, W.J., and G.E.
Noufer, Circuit for address transition detection, US Patent 4563599, 1986.
[11] Huang, J.S.T., and J.W.
Schrankler, Switching characteristics of scaled CMOS circuits at 77K, IEEE
Trans. on Electron Devices ED-34 (1) (1987) 101-106.
[12] Gilchrist, B., J.H.
Pomerene, and S.Y. Wong, Fast Carry Logic for Digital Computers, IRE Trans. on
Electronic Computers EC-4 (4) (1955) 133-136.
[13] Hendrickson, H.C., Fast
High-Accuracy Binary Parallel Addition, IRE Trans. on Electronic Computers
EC-9 (4) (1960) 465-469.
[14] Majerski, S., and M.
Wiweger, NOR-Gate Binary Adder with Carry Completion Detection, IEEE Trans.
on Electronic Computers EC-16 (1) (1967) 90-92.
[15] Reitwiesner, G.W., The
determination of carry propagation length for binary addition, IRE Trans. on
Electronic Computers EC-9 (1) (1960) 35-38.
Appendix
SPICE2G.6: MOSFET model parameters
|
|
|
|
VALUE |
|
Name |
Parameter |
Units |
PMOS |
NMOS |
1 |
level |
model
index |
- |
3 |
3 |
2 |
VTO |
ZERO-BIAS THRESHOLD VOLTAGE |
V |
-1.337 |
1.161 |
3 |
KP |
TRANSCONDUCTANCE
PARAMETER
|
A/V2 |
2.310-5 |
4.610-5 |
4 |
GAMMA |
BULK THRESHOLD PARAMETER |
|
0.501 |
0.354 |
5 |
PHI |
SURFACE POTENTIAL |
V |
0.695 |
0.660 |
6 |
RD |
DRAIN OHMIC RESISTANCE |
OHM |
333 |
85 |
7 |
RS |
SOURCE OHMIC RESISTANCE |
OHM |
333 |
85 |
8 |
CBD |
ZERO-BIAS B-D JUNCTION
CAPACITANCE
|
F |
1.9810-14 |
6.910-15 |
9 |
CBS |
ZERO-BIAS B-S JUNCTION
CAPACITANCE
|
F |
1.9810-14 |
6.910-15 |
10 |
IS |
BULK JUNCTION SATURATION
CURRENT
|
A |
3.4710-15 |
9.2210-15 |
11 |
PB |
BULK JUNCTION POTENTIAL |
V |
0.8 |
0.8 |
12 |
CGSO |
GATE-SOURCE OVERLAP CAPACI-
TANCE PER METER CHANNEL WIDTH
|
F/M |
6.7010-10 |
3.3010-10 |
13 |
CGDO |
GATE-DRAIN OVERLAP CAPACI-
TANCE PER METER CHANNEL WIDTH
|
F/M |
6.7010-10 |
3.3010-10 |
14 |
CGBO |
GATE-BULK OVERLAP CAPACITANCE
PER METER CHANNEL LENGTH
|
F/M |
1.9010-9 |
2.6010-9 |
15 |
RSH |
DRAIN AND SOURCE DIFFUSION
SHEET RESISTANCE
|
OHM/SQ |
55 |
30 |
16 |
CJ |
ZERO-BIAS BULK JUNCTION BOTTOM
CAPACITANCE PER SQ METER OF
JUNCTION AREA
|
F/M2 |
3.5310-4 |
1.2410-4 |
17 |
MJ |
BULK JUNCTION BOTTOM GRADING
COEFFICIENT
|
- |
0.5 |
0.5 |
18 |
CJSW |
ZERO-BIAS BULK JUNCTION SIDE-
WALL CAPACITANCE PER METER OF
JUNCTION PERIMETER
|
F/M |
1.7110-10 |
3.2010-11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|