Professional Documents
Culture Documents
Ddecs 2007 4295258
Ddecs 2007 4295258
Ddecs 2007 4295258
Abstract— Increasing level of process variation in the sub- intra-die or called within-die (∆XW ID ) [6]-[9]. A model can
100nm silicon technology is becoming an important issue. In be expressed as
this paper we describe an approach to estimate the impact
of process variations on the static CMOS and the dual-rail ∆X = ∆XL2L + ∆XW 2W + ∆XD2D + ∆XW ID (1)
PLA down to 32nm process. This approach is built on accurate
variation modeling, published data including the ITRS, Predictive
= ∆XD2D + ∆XW ID−SY S−layout
Technology Models, and Monte-Carlo analysis. The analysis +∆XW ID−SY S−spatial + ∆XW ID−RAN (2)
results show that a challenge due to insufficient noise margins
is posed to the static CMOS at 32nm, and to the dual-rail PLA The first three components in (1) affect all the devices on
from 90nm. Then a one-side virtual ground structure is also the same chip in the same way, e.g., making the transistor gate
proposed to improve the noise margins of the dual-rail PLA. lengths of devices on the same chip all larger or all smaller.
The improved dual-rail PLA is shown to work down to 32nm Thus they are summarized into one term - ∆XD2D
in (2).
process with keeping an operational margin of 150mv.
∆XW ID in (1), on the other hand, may affect different
devices differently on the same chip and they can be further
I. I NTRODUCTION
divided into three components in (2). ∆XW ID−SY S−layout
Process parameter variation is increasing and has been is the systematic layout pattern dependent variation which
expected as a bottleneck in the sub-100nm CMOS technology. is mainly caused by the layout dependency of the process.
It is necessary to analyze, predict and alleviate the impact ∆XW ID−SY S−spatial is the systematic spatial variation. It
of these process variations on integrated circuits for future means the variations of transistors lying near each other are
generations of MOSFET technology. more correlated, than the variations of transistors that are far
Compared with the static CMOS design, array-based design apart. The last component ∆XW ID−RAN , which is caused by
approaches [1]-[3] are supposed to be able to alleviate process process fluctuations, is totally random.
variations and with more manufacturability. In this work, B. Characteristics Analysis of static CMOS and PLA Varia-
we describe an approach to estimate the impact of process tions
variations on these two design styles: the static CMOS and
∆XD2D , ∆XW ID−SY S−spatial and ∆XW ID−RAN are
the PLA (Programmable Logic Array). Our approach is built
substantially independent of the physical implementation of
on accurate variation modeling, published data including the
the IC. Hence, for both the static CMOS and the PLA,
ITRS [4], application of Predictive Technology Models (PTM)
the characteristics of them are the same. ∆XD2D and
[5], and Monte-Carlo analysis to estimate and predict the
∆XW ID−RAN are components of variations with the spa-
trend of noise margins of the two design styles along with
tial correlation of 1 and 0 within a chip, respectively.
technology scaling.
∆XW ID−SY S−spatial is assumed to be systematic with spatial
In Section II, process variations are analyzed in detail. In
correlation=1 on small area in this study [6]-[8].
Section III and IV, the static CMOS and the dual-rail PLA
The difference between the static CMOS and the PLA
are respectively analyzed with respect to process variations.
regarding variations lies in ∆XW ID−SY S−layout . Early in the
In section V, a one-side virtual ground structure is proposed
design stage, the layout of the static CMOS is undecided.
to improve the noise margins of the dual-rail PLA. Section VI
Hence, ∆XW ID−SY S−layout is assumed to be random and
concludes the paper.
thus with spatial correlation ≈ 0. However, the layout of the
PLA is regular and decided; therefore ∆XW ID−SY S−layout
II. VARIATION M ODELS can be assumed to be systematic, with spatial correlation ≈ 1
A. Variation Decomposition and with small magnitude [9][10].
Process variations depend on different scales in time and C. Sources of Variations Considered in This Study
space. In general, these variations can be classified as lot- In this paper we consider process variations in physical
lot (∆XL2L ), wafer-wafer (∆XW 2W ), die-die (∆XD2D ) and gate length (Lg ) and width (Wg ), effective gate length (Lef f ),
9
parallel capacitor across the gate oxide as 8.5
8
ox ox
∆Covs,d = (Lovs,d + ∆Lovs,d ) − Lovs,d 7.5
Tox + ∆Tox Tox 7
6.5
1) Random Fluctuaion of Dopants: In scaled CMOS de- 6
vices, there exists a statistical fluctuation in the number of 5.5
dopants, resulting in a threshold voltage variation (∆Vth ). This 5
discrete dopant effect on ∆Vth is estimated as [11] 4.5
1 2 3 4 5 6 7 8 9 10
Wg/Lg
1 4 si q 3 Nch (|2φf + VBS |) 1
σVth = (5) Fig. 2. Wg /Lg vs. σ(Ion)/mean(Ion) calculated by 1000 runs of Monte-
Cox 8 Lef f Wef f Carlo simulation. Wg /Lg = 4 is chosen for NMOS of the inverter because
when Wg /Lg > 2, the normalized variation of Ion approaches a stable value.
Cox is the gate oxide capacitance per unit area, si is the
silicon permittivity, q is the electron charge, Nch is the channel
doping concentration, φf is the Fermi potential and VBS is Regarding the size of the INV, Wg /Lg = 4 is chosen for the
the body bias voltage, respectively. VBS = 0 is assumed in NMOS based on Fig. 2, which indicates when Wg /Lg > 2,
the following analysis. the normalized variation of the on-current (Ion) of an NMOS
approaches a stabel value. After deciding the size of the
NMOS, the size of the PMOS is chosen to make Vm = V dd/2. 120
LOW noise margin
120
LOW noise margin
HIGH noise margin HIGH noise margin
90nm 65nm Fig. 4. Noise margin profiles of an inverter defined in Fig. 1 along with
1 1 scaling. They are histograms with an interval of 5mv.
0.8 0.8
0.6 0.6 LOW noise margin
0.4 0.4 500
200
45nm 32nm mean 6σ,
σ(Vdd)=50mv
1 1
100
0.8 0.8
0.6 0.6 0
90 65 45 32
0.4 0.4
physical gate length(nm)
0.2 0.2
0 0
0 0.5 1 0 0.5 1 HIGH noise margin
input/output voltage (V) input/output voltage (V) 500
rectangles can no longer be embedded inside the loop when the process shrinks 400
to 32nm.
300
mean 6σ, ideal Vdd
IV. THE VARIATION A NALYSIS OF THE D UAL -R AIL PLA 200
A. The Dual-Rail PLA with Latch Sense Amplifiers [3] mean 6σ,
100
σ(Vdd)=50mv
This section begins with a simple review of the PLA which
will be analyzed in this paper. Fig. 6 shows a column circuit of 0
90 65 45 32
the dynamic dual-rail PLA with latch sense amplifiers (sense-
physical gate length(nm)
amp). Logical OR and NOR of the input signals (X0 − Xn−1 )
can be obtained from the output signals, OU T and OU T ,
respectively. Logical AND is also obtained by performing Fig. 5. The trend of noise margins of an inverter according to the unity gain
definition of Fig. 1 along with scaling. When σV dd = 50mv, there is no
logical NOR of complement input signals. Thus, an AND- HIGH noise margin left for 6σ assurance for an inverter at 32nm. thus we
plane and an OR-plane for a PLA can be realized by arranging predict the static CMOS can not work reliably at 32nm process.
the column circuits. By using a sense-amp, the output signals
are activated by sensing the differential voltage between the
bit-lines, BL and BL. Virtual ground (VG) is provided to Fig. 7 shows a timing diagram of control signals. In phase 1,
reduce the voltage swings of the bit-lines. the P C signal and all the input signals, (X0 −Xn−1 and X0 −
X n-1 loop sense-amp to open loop as Fig. 8 to estimate the input
X n-1 2W Logic Cell
2W
offset voltage (Voff) variations, which is plotted in Fig. 9. Voff
variations increase along with process scaling due to increasing
BL BL process variations.
X0
X0 2W Logic Cell
2W 2W
Logic PC 2W
W Vout+
Part Reference Cell Vout-
W
Sense
VG Vin+ 4W 4W Vin- Amplifier
PC VG Controller
13W
4W 4W SAE
PC Precharge & 6W
Equalization Circuit W=2.25Lg
4W
SAE
4W 4W
Fig. 8. The closed loop sense amplifier used in Fig. 6 is changed to open
2W 2W loop to estimate the input offset voltage variations along with process scaling.
SAE
6W 1 90nm 1 65nm
Vout- (V)
Vout- (V)
0.5 0.5
W=2.25Lg
Vout+ 0 0
Vout+
Fig. 6. The column circuit of the dual-rail PLA proposed in [3] -0.5 -0.5
Vout- (V)
Vout- (V)
BL , BL
0.5 0.5
OUT , OUT
0 0
Vout+
Vout+
-1 -1
Xn−1 ,) are low. Thus, the bit-lines are pre-charged high and -1 0 1 -1 0 1
equalized. At the same time, VG is discharged low. When the Vin+ Vin- (V) Vin+ Vin- (V)
P C signal becomes high and the input signals are activated, Fig. 9. Input offset voltage variations of the sense-amp of Fig. 8
the circuit enters phase 2.
In phase 2, BL is pulled down by charge sharing with To work properly, the differential voltage generated between
VG through a reference cell. When at least one of the basic bit lines of the logic part (Vdiff) must be larger than Voff of
logic cells pulls BL down, the voltage of BL becomes the sense-amp. Fig. 10 shows the variations of Vdiff when
lower than that of BL. Otherwise, BL stays high. This is only one input is high. The upper graph shows the voltage of
because the device size of the basic logic cells is twice of BL and BL. The lower graph shows Vdiff: VBL − VBL . It
the reference cell. The SAE signal is activated when the shows that there is no Vdiff margin from 45nm process.
developed voltage difference between the bit-lines becomes Fig. 11 gives the profiles of Voff and the maximum Vdiff.
larger than the designed sense voltage. By activating the sense- They are histograms with an interval of 5mv. In this paper
amp, one of the output signals, i.e., OU T or OU T becomes the noise margin of the dual-rail PLA is defined as the space
high depending on the developed voltage difference. After the between the Voff and the maximum Vdiff profiles, which is
activation of the sense amplifier, the P C signal becomes low insufficient and disappears from 65nm process.
and the circuit starts to pre-charge the bit-lines.
V. ONE - SIDE VIRTUAL GROUND STRUCTURE (1- SIDE VG)
B. The Variation Analysis of the dual-rail PLA TO IMPROVE THE NOISE MARGINS OF THE DUAL - RAIL PLA
We analyze the trend of noise margins of the described PLA To enlarge the noise margins of the dual-rail PLA, we
above. We divide a column PLA into 2 parts - the logic part propose a one-side virtual ground (1-side VG) structure for
and the sense-amp as shown in Fig. 6. First change the closed the logic part, which is showed as Fig. 12.The voltage of
VBL & VBL (V)
Vdiff (V)
Vdiff (V)
0 0
-0.1 -0.1 X0
0 1n 2n 3n 0 1n 2n 3n X0 2W
Time (s) Time (s) Logic Cell
2W
90nm 65nm PC
W
Reference Cell
VBL & VBL (V)
0.5 0.5 VG
VBL VBL PC
Logic 13W
1 VG Controller
0 0 Part
0 1n 2n 3n 0 1n 2n 3n 4W 4W
Time (s) Time (s) PC Precharge &
0.35 0.35
Equalization Circuit
Vdiff (V)
Vdiff (V)
4W
4W 4W
0 0 SAE
-0.1 -0.1
0 1n 2n 3n 0 1n 2n 3n
Time (s) Time (s)
W=2.25Lg
45nm 32nm BL BL
Fig. 10. Variations of differential voltage between bit lines (Vdiff) when
only one input is high. Of each process, the upper graph shows the voltage of
BL and BL. The lower graph shows Vdiff: VBL − VBL . There is no Vdiff Fig. 12. A one-side virtual ground structure proposed in this paper for the
margin from 45nm process. logic part to enlarge the differential voltage generated between bit lines to
improve the noise margins of the dual-rail PLA (Fig. 6)
VBL & VBL (V)
Vdiff (V)
20 20
0.6 0.6
0
-0.2 -0.1 0 0.1 0.2 0.3 0.4
0
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.4 0.4
voltage(V) voltage(V) 0.2 0.2
0 0
90nm 65nm 0 0.2n 0.4n 0 0.2n 0.4n
Time (s) Time (s)
120 max Vdiff 120 max Vdiff
Voff Voff
90nm 65nm
quantity per 5mV
100 100
VBL & VBL (V)
VBL & VBL (V)
80 80 VBL
1 1
60 60 VBL VBL
0.5 0.5
40 40
0 VBL
20 20 0
0 0.2n 0.4n 0 0.2n 0.4n
0
-0.2 -0.1 0 0.1 0.2 0.3 0.4
0
-0.2 -0.1 0 0.1 0.2 0.3 0.4
Time (s) Time (s)
Vdiff (V)
Vdiff (V)
600 400
Fig. 14. The trend of noise margins of the one-side virtual ground dual-rail Fig. 15. The comparison of noise margins of a static CMOS INV (N MH ),
PLA. Vdiff is measured at 300ps. It shows the noise margins are increased. the original dual-rail PLA and the 1-side VG dual-rail PLA. While there is
This means the improved dual-rail PLA works more reliably along with no noise margin left for the 6σ assurance for the static CMOS from 32nm
process scaling. process and for the original dual-rail PLA from 90nm process, the improved
dual-rail PLA with 1-side VG is shown to work down to 32nm process with
keeping an operational margin of 150mv.