Low-leakage SRAM Design with Dual V
t
Transistors
Behnam Amelifard
University of Southern California
Los Angeles, CA
(213) 740-9481
amelifar@usc.edu
Farzan Fallah
Fujitsu Labs of America
Sunnyvale, CA
(408) 530-4544
farzan@fla.fujitsu.com
Massoud Pedram
University of Southern California
Los Angeles, CA
(213) 740-4458
pedram@ceng.usc.edu
Abstract - This paper presents a method based on
dual threshold voltage assignment to reduce the
leakage power dissipation of SRAMs while maintaining
their performance. The proposed method is based on
the observation that the read and write delays of a
memory cell in an SRAM block depend on the physical
distance of the cell from the sense amplifier and the
decoder. The key idea is thus to realize and deploy
different types of six-transistor SRAM cells
corresponding to different threshold voltage
assignments for individual transistors in the cell. Unlike
other techniques for low-leakage SRAM design, the
proposed technique incurs no area or delay overhead.
In addition, it results only in a slight change in the
SRAM design flow. Finally, it improves the static noise
margin under process variations. Experimental results
show that this technique can reduce the leakage-power
dissipation of a 64Kb SRAM by more than 35%.
1. Introduction
Due to technology scaling, reducing the leakage power
dissipation has become one of the most important
criteria in the design of VLSI systems. The leakage
power dissipation is roughly proportional to the active
area of a circuit. In many processors, caches occupy
about 50% of the chip area [1]. Therefore, the static
power dissipation of a cache is one of the key
components of power dissipation in microprocessors. A
number of researchers have addressed this problem. In
[3], the authors proposed using asymmetric SRAM cells
to reduce the leakage. Their method takes advantage of
the fact that in ordinary programs most of the bits in
data and instruction caches are zero. By including the
device-level optimization into circuit-level techniques,
reference [4] presented a forward body-biasing
technique for active and standby leakage power
reduction in cache memories. In [5], the authors
proposed a dynamic V
t
technique to reduce the leakage
power in SRAMs. In their method, the threshold
voltage of the transistors of each cache line is
controlled separately by using body biasing. In [6], the
use of a diode-connected PMOS bias transistor to
control the virtual ground has been proposed.
Although many techniques have been proposed to
address the problem of low-leakage SRAM design,
most of them result in hardware overhead and hence
increase chip’s area and reduce the manufacturing
yield. In this paper we present a method for low-power
SRAM design based on using different types of cells
with different threshold voltage assignments. The idea
is that due to the non-zero delay of interconnects
different cells in a memory array have different read
and write delays. Therefore, the leakage power
consumption can be reduced by using a high threshold
voltage for some transistors. This technique has four
main advantages over previous techniques:
• It does not have any hardware overhead,
• It does not have any delay overhead,
• It does not need a drastic change in the SRAM
design flow, and
• It improves the static noise margin under
process variation.
The remainder of this paper is organized as follows. In
Section 2 the structure of an SRAM block is
discussed. In Section 3 our idea for reducing the
leakage power dissipation is presented. Section 4
shows the experimental results, while Section 5
concludes the paper.
2. SRAM Design
A typical SRAM block, shown in Figure 1, consists of
cell arrays, address decoders, column multiplexers,
sense amplifiers, I/O, and a control circuitry. The
functionality of the control circuit is to generate
internal signals of the SRAM. In the following, the
functionality and design of other components are
briefly discussed.
A. SRAM Cell
Figure 2 shows a 6-transistor (6T) SRAM cell. The bit
value stored in the cell is preserved as long as the cell
is connected to a supply voltage whose value is greater
than the Data Retention Voltage (DRV) [7]. This
feature, which is due to the presence of cross-coupled
inverters inside the 6T SRAM, holds independent of
the amount of leakage current.
Traditionally all cells used in an SRAM block are
identical (i.e., the transistors with the same name in two
different cells have the same width and threshold
voltage) which results in identical leakage characteristic
for all cells. However, as we shall demonstrate later, by
using non-identical cells, yet with the same layout
footprint, one can realize more power efficient designs.
Cell Array
Ad
d
re
s
sDe
c
od
e
r
Column Multiplexers
Sense Amplifiers
Control
Circuit
I/O
Cell Array
Ad
d
re
s
sDe
c
od
e
r
Column Multiplexers
Sense Amplifiers
Control
Circuit
I/O
Figure 1. An SRAM block.
V
dd
M1
M2
M3
M4
M5
M6
WL
BL
BLB
V
dd
M1
M2
M3
M4
M5
M6
WL
BL
BLB
Figure 2. A 6T SRAM cell.
There are two dominant leakage paths in a 6T SRAM
cell: 1) V
dd
to ground paths inside the SRAM cell and 2)
the bit line to ground paths through the pass transistors.
To reduce the first type of leakage, the threshold
voltages of the pull-down NMOS transistors and/or
pull-up PMOS transistors may be increased, whereas to
lower the second type of leakage, the threshold voltages
of the pull-down NMOS transistors and/or pass
transistors can be increased.
B. Cell Array
Usually there is more than one cell array in an SRAM
circuit. Figure 1 shows only one of them. The size of
the cell array depends on both performance and density
requirements. Generally speaking, as technology
shrinks, cell arrays are moving from tall to wide
structures [2, 13]. However, since using wider arrays
needs more circuitry for column multiplexers and sense
amplifiers, in cases where a large area overhead is
intolerable (e.g., large L3 caches), the number of rows
can be still high [8, 9].
C. Address Decoder
Although the logical function of an address decoder is
very simple, in practice designing it is complicated
because the address decoder needs to interface with the
core array cells and pitch matching with the core array
can be difficult [10]. To overcome the pitch-matching
problem and reduce the effect of wire’s capacitance on
the delay of the decoder, the address decoder is often
broken into two pieces. The first piece, called
predecoder, is placed before the long decoder wires
and the second part, row decoder, which usually
consists of a single NAND gate and buffers for driving
the word-line capacitance, is pitch-matched and placed
next to each row as shown in Figure 3 [10].
D. Column Multiplexers and Sense Amplifiers
Column multiplexing is inevitable in most SRAM
designs because it reduces the number of rows in the
cell array and as a result increases the speed. Since bit
or bit-bar line is discharged about 200mV during a
read operation, a sense amplifier is used to sense a
small voltage difference and generate a digital value.
row1
row2
Predecoder
Row decoder
Cell Array
Decoder wires
A
ddr
e
s
s
Fastest cell
Slowest cell
row1
row2
Predecoder
Row decoder
row1
row2
Predecoder
Row decoder
Cell Array
Decoder wires
A
ddr
e
s
s
Fastest cell
Slowest cell
Figure 3. An SRAM block with its decoder.
3. Hybrid Cell SRAM
Due to the non-zero delay of the interconnects of the
address decoder, word-lines, bit-lines, and the column
multiplexer, read and write delay of cells in an SRAM
block are different. Simulations show that for a typical
SRAM block, the read time of the closest cell to the
address decoder and the column multiplexer is 5-10%
less than that for the furthest cell (cf. Figure 3.) This
gives an opportunity to reduce the leakage power
consumption of the SRAM by increasing the threshold
voltage of some of the transistors in SRAM cells.
On the other hand, due to the delay of sense amplifiers
and output buffers in a read path, the write delay of a
cell is lower than its read delay. Considering the fact
that increasing the threshold voltage of the PMOS
transistors in a 6T SRAM cell increases the write
delay, without having much effect on its read delay,
one may try to reduce the leakage power by increasing
the threshold voltage of the PMOS transistors as long
as the write time is below some target value.
It is known that each additional threshold voltage
needs one more mask layer in the fabrication process,
which results in increasing the fabrication cost [11]. At
the same time, there are studies that show the benefit of
having more than two threshold voltages is small [11].
As a result, in many cases, only two threshold voltages
are utilized. Therefore, we concentrate on the problem
of low-leakage SRAM design when only two threshold
voltages are available. However, it is possible to extend
the results to handle more than two threshold voltages.
A. SRAM Cell Configurations
To reduce the leakage power consumption of a cell, the
threshold voltage of all or some of the transistors of the
cell can be increased. If the threshold voltage of all
transistors within a cell is increased, the leakage
reduction is the highest; however, since this scenario
has the worst effect on the read delay of the cell, the
number of cells that can be replaced is low. Thus, we
consider other configurations that have smaller leakage
reductions due to their lower delay overheads.
Unlike [3], we use a symmetric cell configuration,
which means the symmetric transistors within a cell
have the same threshold voltages. Thus, there are eight
different possibilities for assigning high and low
threshold voltages to the transistors within a cell. These
configurations are shown in Table 1. Here it is assumed
that the threshold voltage of each transistor in the
SRAM cell can be adjusted independent of other
threshold voltages by changing the channel doping,
which is deemed to be a safe assumption because in an
SRAM cell the channels of the transistors are not too
close to each other [14]. However, in the Simulation
Results Section it will be shown that even if only one
threshold voltage is used inside the cell, power
reduction can be considerable.
The configuration ordering is such that, excluding C0
which is the original configuration, the leakage current
saving of other configurations is monotonically
decreasing as we move from C1 toward C7. Figure 4
shows the leakage current saving of different
configurations versus the value of the high threshold
voltage. The numbers in Figure 4 and the following
ones are obtained using a 180nm CMOS technology
with 1.8V for the supply voltage and 0.37V for the low
threshold voltage at 100
0
C.
Each of these configurations has different effects on
read and write delays of cells. Figures 5 and 6 show the
increase in read and write delays for each configuration
for different values of the high threshold voltage. It can
be seen that the increase in read delay for some
configurations (e.g., C1 and C3) is very high, whereas it
is very small for some other configurations (e.g., C6). It
can also be seen from Figure 6 that not all
configurations increase the write time; In fact, C4 and
C7 decrease the write time.
Table 1. Possible configurations for high threshold
voltage assignment
Config. High threshold transistors
C0 None
C1 M1, M2, M3, M4, M5, M6
C2 M3, M4, M5, M6
C3 M1, M2, M5, M6
C4 M1, M2, M3, M4
C5 M5, M6
C6 M3, M4
C7 M1, M2
0
10
20
30
40
50
60
70
80
90
100
0.37 0.4 0.43 0.46 0.49 0.52 0.55 0.58
V
t ,high
Leakage current
reduction (%)
C1
C2
C3
C4
C5
C6
C7
Figure 4: Leakage current reduction for each
configuration.
0
10
20
30
40
50
60
0.37 0.4 0.43 0.46 0.49 0.52
V
t,high
Read delay increase (%)
C1
C2
C3
C4
C5
C6
C7
Figure 5: Read delay increase for each configuration.
B. Noise Margin
The static noise margin (SNM) of a CMOS SRAM
cell is defined as the minimum DC noise voltage
necessary to flip the state of a cell [
12]. SRAM cells
are especially sensitive to noise during a read
operation because the “0” storage node rises to a
voltage higher than ground due to a voltage division
along the pull-down NMOS transistor and the pass
transistor; if this voltage is high enough, it can change
the cell’s value. In general, it is expected that by using
high threshold transistors in the SRAM cells, the static
noise margin would increase. We measure the SNM of
each configuration under two scenarios: nominal
condition and process variation.
B.1 Static Noise Margin under Nominal
Conditions
Simulation results shown in Figure 7 confirm that for
all configurations except C6 (i.e., when only PMOS
transistors are high threshold), the nominal SNM is
more than that of C0 and improves with increasing the
high threshold voltage. In C6, however, the SNM is
slightly less than that of C0 and degrades with
increasing the high threshold voltage.
-1
0
1
2
3
4
5
6
0.37 0.4 0.43 0.46 0.49 0.52
V
t,high
Write delay increase (%)
C1
C2
C3
C4
C5
C6
C7
Figure 6: Write delay increase for each Configuration.
B.2 Static Noise Margin under Process Variation
Technology scaling has made the process variation one
of the concerns of designers. As the minimum size
transistors are typically used in SRAM cells to achieve
a compact design [2], SRAMs are very sensitive to
process variation. To measure the static noise margin of
each configuration under process variation, we used the
Monte Carlo simulation technique. For each
configuration we used 400 vectors of six threshold
voltages, where each threshold voltage has a Gaussian
distribution with its mean equal to the nominal value
and its 3σ/µ equal to 0.15.
The simulations show that the distribution of SNM of
each configuration under process variation can be
approximated by a Gaussian distribution. The mean and
standard deviation of SNM for different configurations
are shown in Table 2. Here the important value is µ-3σ
and as one can see in the table, all configurations have a
better µ-3σ than C0.
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
0.38 0.42 0.46 0.5 0.54 0.58
V
t,high
Normalized SNM
C0
C1
C2
C3
C4
C5
C6
C7
Figure 7: SNM for different configurations under the
nominal condition.
Table 2. Mean and standard deviation of SNM for
different configurations (V
t,high
=0.47V).
Config.
µ σ µ-3σ
C0 0.255 0.0217 0.190
C1 0.312 0.0235 0.242
C2 0.287 0.0224 0.220
C3 0.318 0.0187 0.262
C4 0.269 0.0175 0.217
C5 0.293 0.0165 0.244
C6 0.245 0.0143 0.202
C7 0.276 0.0268 0.196
C. Hybrid Cell Assignment
To design a hybrid-cell SRAM, we need to find out
the slowest read and write delay starting with all low-
V
t
SRAM cells (C0 case.) Next, since C1 results in the
highest leakage reduction among all configurations,
we replace as many C0 cells as possible with C1 cells
in such a way that the access delay of the replaced
cells will not be larger than the slowest access delay.
After that, we try to replace the remaining C0 cells
with C2, C3, C4, C5, C6, and C7.
Figure 8 shows the pseudo-code of the hybrid cell
assignment. rownum and colnum are the number of
rows and columns of the SRAM, respectively.
Moreover, v
tH_lb
and v
tH_ub
are the lower and upper
bounds of the high V
t
value, respectively. The fastest
cell is denoted by index [0, 0], while the slowest one is
denoted by index [colnum-1, rownum-1]. Subroutines
ReadDelay(col,.row,.config) and
WriteDelay(col,.row,.config) return the read and write
delays of cell [col,.row] when configuration config is
used. If cell [col,.row] fails working with
configuration config, then all cells [i, j], where i
≥
.col
and j
≥
.row fail with the same configuration. Therefore,
a large number of cells can be pruned as soon as a cell
fails working for a given configuration. In the pseudo-
code, flag[config][col, row] is a flag that specifies if
cell[col,.row] can work with configuration config.
Initially all flags are set to 1. In the next section, it will
be shown that this algorithm, despite its simplicity,
results in a significant power reduction in SRAM
blocks.
To speed up the process, instead of checking for
possible replacement on each single SRAM cell, one
can select n
×
n blocks and do the checking for the
slowest cell in the block. If the slowest cell passed the
delay test, the whole block is replaced; otherwise, next
configuration or block is examined. Here n is a multiple
of two and it is clear that choosing a larger number for
n decreases the design time, but degrades the result.
Hybrid-Cell-Assignment (rownum, colnum, v
tH_lb
, v
tH_ub
)
Begin
1. T
max
=ReadDelay (colnum-1, rownum-1, C0)
2. For v
t,high
= v
tH_lb
to v
tH_ub
3. For config=C1 to C7
4. For (0≤col<colnum, 0≤row<rownum)
5. flag[config][col,row]=1;
6. For col=0 to colnum-1
7. For row=0 to rownum-1
8. For config=C1 to C7
9. If (flag[config][col,row] ==1)
10. If (ReadDelay(col,row,config)<T
max
&& WriteDelay(col,row,config)<T
max
)
11. Replace cell[col][row] with config; Break;
12. Else
13. For (i≥col, j≥row)
14. flag[config][i,j]=0;
End
Figure 8. Pseudo-code for the hybrid cell assignment.
4. Simulation Results
To study the efficiency of the proposed technique, a
400MHz, 64Kb SRAM with a 64-bit word has been
designed and simulated using Cadence UltraSim in
180nm CMOS technology with 1.8V for the supply
voltage and 0.37V for the low threshold voltage. The
SRAM consists of two 256×128 cell arrays. For
optimizing the delay of the decoder, the predecoding
scheme has been used as described in Section III. The
simulations results in this section are pre-layout.
However, by modeling all local and global
interconnects, including bit and bit-bar lines, word line,
and decoder wires as distributed RC circuits, the
accuracy of the simulations has been improved.
Table 3 shows the leakage power reduction achieved
and the utilization of each configuration in the low-
leakage SRAM for different values of the high
threshold voltage. One interesting observation from
Table 3 is that the leakage saving is not very sensitive
to the threshold voltage value if the value is larger than
0.42V.
From Table 3 it can be seen that C3 is rarely used in the
low-leakage SRAM. The reason is that the delay of this
configuration is almost equal to the delay of C1, but its
leakage is higher. Therefore, in most cases, C1 is used
instead of C3. Similarly, from Figure 5 one can see
that the increase in the read delay of C5 and C7 are
almost equal to those of C2 and C4, respectively; so it
is not surprising that in Table 3 most of the entries for
C5 and C7 are zero. Figure 9 illustrates the
approximate floor-planning of each type of cell for
two values of the high threshold voltage (leakier areas
are shown in darker gray.)
Table 3. The power reduction and the utilization of each
configuration in the low-leakage SRAM.
Utilization of Each Configuration
(%)
V
t,high
Leakage
Reduction (%)
C0 C1 C2 C3 C4 C5 C6 C7
0.38
14.64
13.5 76.2 3.1 0.8 2.1 1.2 3.1 0
0.39
25.00
13.7 65.4 7.6 0 6.4 0 6.8 0
0.40
31.84
13.7 52.9 10.0 0 16.0 0 7.4 0
0.41
36.48
13.7 46.9 8.2 0 18.6 0 12.7 0
0.42
38.77
14.1 33.6 15.0 0 24.0 0 13.3 0
0.43
36.55
14.1 8.0 33.8 0 25.4 0 18.8 0
0.44
34.74
14.1 0.6 22.1 0.2 38.1 0 25.0 0
0.45
35.06
14.1 0 9.4 0.2 51.0 0 25.4 0
0.46
34.66
14.1 0 2.5 0 52.5 0 30.9 0
0.47
35.30
14.1 0 0 0 53.9 0 32.0 0
0.48
35.14
14.1 0 0 0 49.0 0 36.9 0
0.49
34.54
14.1 0 0 0 43.0 0 43.0 0
0.50
33.61
15.8 0 0 0 38.7 0 45.5 0
0.51
33.36
16.2 0 0 0 36.1 0 47.7 0
By observing the fact that C3, C5, and C7 are
dominated by other configurations, they may be
deleted from the set of candidate configurations,
resulting in five configurations: C0, C1, C2, C4, and
C6. If, for any reason, a designer likes to use a smaller
number of configurations, the saving in the leakage
will decrease. We have repeated our experiment for
the case that only three configurations are utilized. In
this case, the selection of suitable configurations
depends on the value of V
t,high
; for example, from
Table 3 it can be seen that when V
t,high
is close to the
low threshold voltage, the best configurations are C0,
C1 and C2, while for higher value of V
t,high
, C0, C4
and C6 ought to be used. Table 4 shows the
configurations used in each case.
Table 5 reports the power reduction of the SRAM
block for different values of the high threshold voltage
when only three different configurations are utilized.
From this Table it can be seen that even when using
only three configurations, the power reduction is high.
In fcat, in most cases the power saving is more than
30%. Table 5 also reports power reduction of the
SRAM block when all threshold voltages in an SRAM