Académique Documents
Professionnel Documents
Culture Documents
art ic l e i nf o
a b s t r a c t
Article history:
Received 18 May 2014
Received in revised form
25 November 2014
Accepted 12 January 2015
Available online 21 January 2015
In this paper, a new write assist technique for SRAM arrays is proposed. In this technique, to improve the
write features of the SRAM cell, a negative voltage is applied to one of the bitlines in the SRAM cell while
another bitline is connected to a boosted voltage. Improved write features are attributed to the boosting
scheme from both sides of the SRAM cell. This technique is applied to a 10T-SRAM cell with
transmission-gate access devices. The proposed design gives 2.7 , 2.1 faster write time, 82% and
18% improvement in write margin compared with the standard 8T-SRAM cell with and without write
assist, respectively. All simulations have been done in TSMC 65 nm CMOS technology. The proposed
write assist technique enables 10T-SRAM cell to operate with 24% lower supply voltage compared with
standard 8T-SRAM cell with negative bitline write assist. Due to the improved supply voltage scalability a
33% leakage power reduction is achieved.
& 2015 Elsevier B.V. All rights reserved.
Keywords:
SRAM
Write margin
Read static noise margin
Write assist
1. Introduction
CMOS technology has been the cornerstone of semiconductor
devices for years. Moore's law motivates the technology scaling in
order to improve the performance features such as speed, power
consumption and area. Although circuits and systems benet from
technology down-scaling in some aspects, the undesired features
such as short channel effects (SCEs) and sensitivity to process
variations are also consequential. The effect of process variations on
performance is a key issue in scaled CMOS technology. This effect
gets more pronounced as the size of transistors is reduced. One of the
highly sensitive circuits to process variations is SRAM due to using
extremely small devices in order to achieve high density. Process
variations can be due to global or local mismatches between devices.
Global mismatch refers to die-to-die variation in device and local
mismatch refers to mismatch between transistors on the same die
[1]. Local mismatch in SRAM devices can easily lead to read stability
degradation (stored data is ipped during read), read failure (data is
not read during read period), writeability decrease or write time
increase. On the other hand, the trade-off between read and write
operations makes it difcult to improve both simultaneously.
Fig. 1(a) shows the standard 6T-SRAM cell structure. It consists
of two back to back inverters (PUL-PDL and PUR-PDR) that keep
n
Corresponding author.
E-mail addresses: Farkhani.hooman@stu.um.ac.ir (H. Farkhani),
Peiravi@um.ac.ir (A. Peiravi), moradi@eng.au.dk (F. Moradi).
http://dx.doi.org/10.1016/j.vlsi.2015.01.001
0167-9260/& 2015 Elsevier B.V. All rights reserved.
the data and its inverse on nodes Q and QB, respectively. The
access transistors (ACL-ACR) are used to perform read and write
operations. Due to using a common path (ACL-ACR) for read and
write operations, improving the read stability will lead to degradation of the writeability of the cell and vice versa. To improve the
read stability of an SRAM cell, beta ratio ( WPD/WAC) can be
increased, while lower alpha ratio ( WPU/WAC) is desirable to
improve the cell writeability. Finally, during hold mode, equal
strength for pull up and pull down transistors (WPU WPD) sets the
trip point of the two back-to-back inverters at VSUPPLY/2 and
ensures maximum noise margin.
Several solutions have been proposed in literature from device to
architecture level to improve SRAM cell functionality. For instance,
at the device level, using new devices such as FinFET, leads to
signicant SRAM performance improvement [15]. At the cell level,
new cells such as 7T, 8T, 9T, 10T, and 11T [515] have been proposed
that have resulted in a higher cell area. Among them standard 8TSRAM cell shows a very good compromise between cell features
improvement and cell area penalty [6]. Fig. 1(b) shows the standard
8T-SRAM cell. This cell consists of a 6T storage cell and a 2T read
buffer which isolates read path from write path and storage nodes.
In this way, read and write features can improve simultaneously.
However, the write margin of the SRAM cell varies signicantly due
to the process variations of access transistors. This is due to the fact
that the write current ows through the access transistors and
corresponding variations will affect the write margin of the SRAM
cell signicantly. On the other hand, while reading the cell, the read
17
Fig. 1. (a) Standard 6T- (b) Standard 8T- (c) PPN 10T- [13] (d) 10T- [14] (e) TGA 10T-SRAM cells [15].
current ows through the read buffer (R1 and R2) which its
variation does not affect read margin. As a result, the write margin
is a key parameter especially in the scaled supply voltages which
the effect of process variation is more pronounced. Supply voltage
scaling is an effective way to decrease the power consumption of
the System-on-Chips (SoCs). 10T-SRAM cells are mainly proposed to
work in ultra-low voltages (near/subthreshold region). Fig. 1(c)
(e) shows three 10T-SRAM cell schematics. Lo et al. proposed a PP
N 10T-SRAM cell in [13]. This cell is called PPN due to the fact that
each of the cross-coupled inverters consists of three transistors
cascaded in a PPN sequence as shown in Fig. 1(c). This cell shows
a low cell leakage due to the three stacked transistors as inverter
and it also has a good immunity to data dependent bitline leakage
[13]. However, the use of stacked transistors limits supply voltage scaling. Fig. 1(d) shows another 10T-SRAM cell in the literature
which uses two transistors in series as access device [14]. Utilizing
two control signals (WL and W_WL) enables implementation of bitinterleaving structure. The main drawback of this cell is related to its
weak writeability due to the cascaded access transistors which can
limit the power supply scaling. Fig. 1(e) shows the TransmissionGate Access transistor SRAM cell (TGA-SRAM) proposed in [15]. The
transmission gate access devices are used to improve the writeability of the cell that improves the supply voltage scalability, as
well. On the other hand, the read path is isolated from storage nodes
that results in a high read noise margin. Single ended read sensing
scheme will decrease the TGA-SRAM cell access time. However,
utilizing pseudo-differential sensing improves the access time,
signicantly [16]. All in all, each of the 10T-SRAM cells has some
pros and cons which mentioned above. However, comparing the
cells is not the focus of this paper and we will use only one of the
above mentioned cells in order to apply the proposed write assist
technique.
At the architecture level, read and write assist techniques
improve SRAM robustness and performance while occupying less
area compared to the cell techniques (e.g. 8T and 10T) and can be
used with any type of SRAM.
In this paper, a new write assist technique is proposed that
improves SRAM writeability, signicantly. This new technique
applies a boosted voltage (VDD V) to the high-going bitline
and a negative voltage on the low-going bitline simultaneously.
This technique results in a signicant improvement in Write
18
Hold Static Noise Margin (HSNM) of the half-selected cells and are
not suitable for low voltage applications.
Among the existing techniques, negative bitline write assist is
the most effective one to reduce SRAM minimum supply voltage,
especially when considering dynamic failure metrics [27].
Fig. 2. Use of a replica column along with write buffer in order to make a negative
bitline voltage during write operation [25].
Fig. 3. Use of a replica column with a replica write buffer in order to nd the best time of connecting the boost capacitor [21].
19
Fig. 4. Using two boosted capacitors in order to make a negative voltage on the
appropriate bitline [23].
BL
BLB
WL
WL
Q=0'
QB=1'
ACR
ACL
SACL
SACR
BL
BLB
WL
TNL
WL
TNR
Q=0
1
TPL
WLB
QB=1
RBL
RWL
0
TPR
WLB
R2
R1
STNL-TPL
STNR-TPR
Fig. 5. Write operation on (a) SRAM cells with conventional write path and
(b) TGA-SRAM cell [14].
20
RWL[0]
WLB[0]
RBL[m]
BLB[m]
BL[m]
RBL[0]
CELL
BLB[0]
BL[0]
t0
WL[0]
t1
WR
RBL
RP
RWL[n]
VDD+
WL[n]
BL
CELL
WLB[n]
BLB
BLS[0-m]
Column Decoder
BLBS[0-m]
REBL
Sense Amp.
C1
WR
C2 C1
C1 C2
C2
C1 C2
DB
MP
CP-boost
+
WD1
VB
WD2
CN-boost
-
WR
VN
RP
Replica Column
MN
Data
RP
C2
WR
RP
D
Data
C1
DB
RPUR
ITNR
ITNR
CN
VQB
VQB
CQB
!
a2 t
p
p
2
b2 j b2 4a2 c2
j
b2 2 4a2 c2
CQ
p
2a 1
e
jb2 j b2 2 4a2 c2
VQ
VBLB
VN
VBLB
ITPR
ITPR
CBLB
where
a2
TNL
2
1
b2 TNL V DD V THN
RPDL
TGA_SRAM Cell
WLB
21
c2
TNL
2
V DD V THN 2
TPL
2
jV BL j jV THP j2
WL
RPDL
Fig. 8. (a) The equivalent simple circuit of TGA-SRAM cell at the side of QB node
and (b) voltage variation of Q and QB while write operation in SRAM cell with
conventional write path and TGA-SRAM cell.
respectively. In order to see the effect of BNBL write assist on TGASRAM cell writeability, we nd equations for VQB and VQ. To simplify
the equations, VBL and VBLB are assumed to be constant. When the
write cycle starts, PUR is in its linear region and TNR and TPR are in
saturation region. A KCL at node QB gives:
C QB
dV QB
I PUR I TNR I TPR
dt
I PUR
I TNR
I TPR
V DD V QB
RPUR
TNR
2
TPR
2
V DD V BLB V thn 2
V QB j V thp j 2
a1 t
p
q
q
2
2
2
b1 b1 4a1 c1 Ze CQB b1 4a1 c1 b1 b1 4a1 c1
!
a1 t
p
2
Ze CQB b1 4a1 c1 1 j2a1 j
PD V QB V THN
Eq. (6) shows the voltage equation for the node Q holding 0
while writing 1 on it in TGA-SRAM cell. The parameters a2 and c2
are positive while b2 is negative. By boosting the voltage on BL using
the proposed write assist technique, c2 value will increase. In order
to explore the effect of growth in c2 on VQ, we simplied Eq. (6) in
(7) and the increment and decrement in each parenthesis are
shown withand , respectively. As can be seen in (7), increase of
c2 leads to an increase in both numerator and denominator of
Eq. (7). However, the pace of increase in the numerator is much
faster than the denominator that causes a faster rise in VQ. As a
result, it improves write margin and write time of the proposed cell.
Faster increase of VQ has two positive effects on the write operation
as it can be seen in Fig. 8(b). First, faster increase of VQ leads to a
much shorter write time. This is due to the fact that contrary to that
of conventional SRAM cells, where charging the VQ is mainly done
by the back to back inverters, in TGA-SRAM cell, charging node Q is
done through both access devices connected to the boosted bitline
and back to back inverters. The second reason is that faster increase
in VQ means faster decrease of VSG-PUR (source-gate voltage of TPUR)
that will lead to a faster decrease in the strength of TPUR. Therefore,
access devices (TNR-TPR) can discharge the storage node QB easier
(Fig. 8(b)). Faster discharge of node QB has the same effect on node
Q and it works like a positive feedback that improves write margin
and write time, signicantly.
VQ
1
2a 1
5
where a1 , b1 , c1 , and Z are:
a1
TPR
2
b1 TPR jV THP j
c1
4. Simulation results
1
RPUR
V DD TPR
q
2a V b b 2 4a c
1
1
1 1
1 DD
q
Z
2
2a1 V DD b1 b1 4a1 c1
RPUR
PU V DD V Q jV THP j
22
Table 1
Transistors width of standard 8T- and TGA-SRAM cells.
PUR (nm)
PDR (nm)
R1 (nm)
R2 (nm)
ACL (nm)
ACR (nm)
TNL (nm)
TPL (nm)
TNR (nm)
TPR (nm)
150
150
150
150
150
150
150
150
180
180
180
180
180
180
180
230
180
230
PDL (nm)
Standard 8T
TGA
PUL (nm)
1.2
1
Q node
QB node
0.8
0.6
0.4
Write Margin
0.2
0
-0.2
0.2
0.4
0.6
0.8
1.2
1
Q node
QB node
0.8
0.6
238mV
0.4
0.2
0
-0.2
0.2
1.2
1
Q node
QB node
0.8
0.6
186mV
0.4
0.2
0
-0.2
0
0.2
0.4
0.6
0.8
Q node
QB node
0.6
162mV
0.4
0.2
0
-0.2
0
Q node
QB node
82mV
0.2
0
-0.2
0.2
0.4
0.6
WL voltage (V)
0.2
0.4
0.6
0.8
1.2
WL voltage (V)
0.8
0.8
1
0.8
0.4
0.6
1.2
WL voltage (V)
0.6
0.4
WL voltage (V)
WL voltage (V)
0.8
1.2
1
Q node
QB node
0.8
80mV
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
WL voltage (V)
Fig. 9. Write margin (VDD minus ipping point voltage) of the TGA-SRAM cell when utilizing (a) no write assist, (b) proposed BNBL technique with VBL 1.2 V and
VBLB 0.2 V, (c) NBL technique with VBLB 0.2 V, (d) boosted WL technique with VWL 1.2 V, (e) cell VDD collapse technique with VDD 0.8 V, (f) cell GND boost with
VGND 0.2 V. The difference between the dashed line and the ipping point of the storage nodes shows the write margin improvement.
23
t0
t1
Case 1:
WL enabled after discharging
the low-going bitline to ground
1.5
WL
VDD+
Case 2:
WL enabled at the start of
write cycle
1
0.5
BL
BLB
-
0,1
-0.1 , 1.1
-0.2 , 1.2
-0.3 , 1.3
1.2
the difference between VDD and WL voltage when the data stored
in the cell is ipped.
Fig. 9 shows the write margin for the TGA-SRAM cell when
utilizing: (1) no write assist technique; (2) proposed BNBL technique (VBL 1.2 V and VBLB 0.2 V); (3) NBL technique (VBLB
0.2 V); (4) BWL technique (VWL 1.2 V); (5) cell VDD collapse
(VDD 0.8 V), and; (6) cell GND boost (VGND 0.2 V) techniques. A
20% voltage increase for boosted BL, cell GND boost and boosted
WL techniques is considered while a 20% voltage decrease is used
for cell VDD collapse and negative BL techniques in order to make a
fair comparison.
The WM of the proposed BNBL technique (2) is higher than the
existing techniques (36). The WM improvements of BNLB, NBL, BWL,
cell VDD collapse and cell GND boost techniques over the cell without
write assist technique are 54%, 42%, 36%, 19% and 18%, respectivly. The
BNBL technique has the highest WM improvement. This is attributed
to the assisted write operation from both sides of the cell. Cell GND
boost and cell VDD collapse techniques have the minimum WM
improvement. This is due to the fact that boosting cell GND decreases
VDS of the access devices connected to Q and increases the trip point
of the right hand side inverter (Fig. 1(e)). On the other hand, cell VDD
collapse technique decreases VDS of the access devices connected to
QB and decreases the trip point of the right hand side inverter.
The NBL write assist technique is the best technique among the
existing techniques in order to improve the WM. In the rest of the
paper, the effect of proposed BNBL technique on TGA-SRAM cell is
compared with TGA-SRAM cell without write assist and 8T-SRAM
cell with NBL write assist technique.
4.2. Write margin
Fig. 10 shows the normalized WM for different negative and
boosted bitline voltages. WM is normalized to the WM of TGASRAM cell without write assist, 8T-SRAM cell with NBL write assist
and without write assist. Fig. 10 justies the idea that applying
simultaneous forces to both sides of the cell improves its writeability.
As illustrated in Fig. 10, increasing the VBBL and |VNBL| improves WM of
TGA-SRAM cell compared with 8T-SRAM cell without write assist
signicantly (2.2 improvement at VBBL 1.3 V and VNBL 0.3 V).
On the other hand, applying BNBL technique improves WM by 18%
compared with 8T-SRAM SRAM cell with NBL write assist. It should be
noted that higher VBBL and |VNBL| voltages needs larger boosting
capacitors which increase the area overhead. Besides, it can cause
higher failure probability in unselected cells by turning on their access
transistors. In order to deal with this issue, we chose 0.2 V and 1.2 V
voltage levels for negative and boosted bitlines, respectively. In this
way the gate-source voltage of access transistors in unselected cells
are far enough from their threshold voltage and they will be OFF
Fig. 10. Write margin of the TGA-SRAM cell while using proposed write assist
technique. Write margin normalized to the write margin of TGASRAM cell without
write assist and standard 8T-SRAM cell with and without NBL write assist.
0.8
0.6
0.4
0.2
0
-0.1 , 1.1
0,1
-0.2 , 1.2
-0.3 , 1.3
during write cycle. Besides, the negative and boosted voltages are
sufcient enough to make a large improvement in WM. For VBL and
VBLB at 0.2 V and 1.2 V, WM is improved by 54%, 18% and 82%
compared with TGA-SRAM cell without write assist, standard 8TSRAM cell with and without NBL write assist, respectively.
4.3. Write time
In order to nd the write time, the time between asserting WL
and when both storage nodes reach to 90% of their nal values is
measured. Fig. 11(a) shows the timing diagram during write cycle
where t0 is the beginning of write cycle and t1 is when low-going
bitline is discharged to ground. The write operation can be divided
into two parts: (a) Discharging the appropriate bitline to ground
(t0 ot ot1). (b) Applying the write assist technique to the bitlines
(t4 t1). As illustrated in Fig. 11(a), during write cycle, WL signal
can be asserted from the beginning of write operation (case 1) or
after low-going bitline is discharged to ground (case 2). However,
in NBL write assist technique, the coupling capacitor needs to be
connected to the low-going bitline after it is fully discharged to
ground, otherwise the negative voltage will be higher and WM
improvement decreases drastically [21]. Fig. 11(b) shows the effect
of the proposed write assist technique on the write-time. As
illustrated in Fig. 11(b), applying a 0.2 V negative voltage and
1.2 V boosted voltage to the bitlines leads to 2.7 and 1.5
improvement in the write time compared to the 8T-and TGASRAM cells without write assist circuit. Besides, the proposed
write assist technique results in 2.1 improvement in write time
compared with the 8T-SRAM cell using NBL write assist circuit.
4.4. Half-selected cells
In this section, the effect of the proposed write assist technique
on the half-selected cells of TGA-SRAM cell is explored.
24
Voltage (mV)
600
500
400
300
200
100
0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
200
-6 (mV)
150
100
50
WM-TGA
0
0
0.1
0.2
0.3
0.4
0.5
-50
65mV
-100
135mV
Supply Voltage (V)
Fig. 12. (a) Write margin, half-column selected HSNM and RSNM vs. supply voltage
(b) 6-worst case vs. supply voltage for 8T-SRAM cell with NBL, TGA-SRAM cell
with BNBL write assist, and without write assist. Note that the negative points on
this gure are averages over 1000 runs that indicate error in 6-worst case of
related parameters.
10000
1.22
8T with NBL
TGA
TGA with BNBL
1000
25
100
10
80
1.21
60
1.2
40
20
1.19
500
1000
0
1.188 1.198 1.208 1.218 1.228
Iteration number
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
-0.18
Fig. 13. Write time versus supply voltage for 8T-SRAM cell with NBL, TGA-SRAM
cell with BNBL write assist, and without write assist.
-0.192
80
-0.194
60
60
-0.2
40
-0.21
20
-0.22
-0.23
40
500
Iteration number
-0.198
1000
0
-0.228 -0.218 -0.208 -0.198 -0.188
Fig. 15. The effect of process variation on (a) negative voltage level and (b) boosted
voltage level
20
-0.2
0
500
1000
0
-0.204
Iteration number
80
-0.19
-0.196
0.2
Negative BL voltage (V)
-0.2
-0.196 -0.192
1.204
80
1.202
60
1.2
40
1.198
20
1.196
0
500
Iteration number
1000
1.194
1.198
1.202
1.206
SRAM array without write assist technique has 1.14x lower DWP.
However, considering the signicant leakage power consumption
improvement (1.48 and 2.47 ) of TGA-SRAM array with BNBL
technique, it has the minimum total power consumption.
Fig. 13 shows the write time of 8T-SRAM cell with NBL, TGASRAM cell with BNBL and TGA-SRAM cell without write assist
circuit at different supply voltages. The write time of TGA-SRAM cell
with BNBL write assist technique is signicantly shorter than 8TSRAM cell with NBL and TGA-SRAM cell without write assist
technique at all simulated supply voltages (e.g. 2.11 and 4.8
more than 8T with NBL and TGA without write assist at 0.4 V supply
voltage). This is attributed to turned-on strengthened access transmission gate devices in TGA-SRAM cell during whole write cycle.
Fig. 14. The effect of process variation on (a) negative voltage level and (b) boosted
voltage level.
26
5. Conclusion
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
Hooman Farkhani received the B.Sc. degree in electrical engineering from Kashan University, Kashan, Iran,
in 2004. He received the M.S. and Ph.D degrees in
electrical engineering from Ferdowsi University of
Mashhad, Mashhad, Iran, in 2008 and 2015, respectively. He is a research assistant in Integrated Circuit
Design (ICD) laboratory in Ferdowsi University of
Mashhad. He visited the ICE-LAB, Aarhus University,
Aarhus, Denmark from August 2013 to November 2013
and currently collaborates with ICE-LAB on beyondCMOS circuit and memory design. He is working on
spin-transfer torque random access memory (STTRAM). His other elds of interest are low power and
low voltage SRAM design and ADCs.
27