1

A 256kb Sub-threshold SRAM in 65nm CMOS

Benton H. Calhoun, Anantha Chandrakasan

Massachusetts Institute of Technology, Cambridge, MA

ISSCC 2006 / SESSION 34 / SRAM / 34.4

Presented by: Pouya Kamalinejad

2006/12/28

Advanced VLSI class presentation

2

Outline

• Introduction and preliminaries.

• SNM introduction.

• Proposed 10T SRAM

• Simulation and results

• Read SNM free SRAM

• Conclusion

3

Why Low Voltage SRAMs?

• The minimum supply voltage of LSIs is limited by their SRAMs for the following two reasons[2]:

1) with decreasing supply voltage (Vdd), SRAM delay increases at a higher rate than does CMOS logic circuit delay.

2) Read operations at low-Vdd levels result in storage data destruction in SRAM cells.

4

Preliminaries

Traditional 6-T SRAM column[3] Bitline discharging for the reaoperation[3]

5

Static Noise Margin

• The large fraction of chip area often devoted to SRAM makes low power SRAM design very important.

• SNM quantifies the amount of voltage noise required at the internal nodes of a bitcell to flip the cell’s contents.

• degraded SNM can limit voltage scaling for SRAM designs.

Inverter 2Inverter 1

SNM is length of side of the largest embedded square on the butterfly curve

WL BLBBL

Q QBM1

M2

M3

M4M5

M6

VN

VN

[1]

6

Cont’d

Moves upward

Moves to the left

SNM is lower during read access because the VTC is degraded by the voltage divider across the access transistor (M2,M5) and drive transistor (M1,M4)[2]

SNM Butterfly Curve

The minimum supply voltage of SRAMs is determined by both Read SNM and Write SNM levels; reducing Vth in the NMOS transistor improves Write SNM but worsens Read SNM.

7

SNM during HOLD and READ

Read SNM is worst-case

WL=0BLBBL

1 0M1

M2

M3

M4M5

M6

WL=1BLB prech 1BL prech 1

1 0M1

M2

M3

M4M5

M6

[1]

8

Sub-VT SNM Dependencies

• SNM is mainly a function of:

Vdd (limited to Vdd/2)Temperature (higher temp results in

Lower SNM due to lower gain)

Sizing (Cell ratio affects SNM less in

sub-threshold due to logarithmic relation

unless it affects Vt)

Bit-line voltageVt mismatch

Vt mismatch is the worstNormal distribution

Model* gives good estimate forthe distribution of SNM at the

worst-case tail

[1]

9

How to reduce Vdd?

Impact of local mismatch on 6T SNM in 65nm. Read SNM has larger standard deviation. Hold SNM at 0.3V has roughly the same mean as Read SNM at 0.5V and same 6σ SNM as Read SNM at 0.6V.[2]

Thus, by eliminating the degraded Read SNM, a bitcell can be operated at 0.3V with the same 6σ

stability as a 6T bitcell at 0.6V. A

10

Cont’d

• The idea is to add a 4T buffer at one side:

6T bitcell

4T buffer

WLBLBBL

Q QB

VVDD

RBLRWL

Proposed 10-T bitcell for Sub-VT[1]Proposed 10-T bitcell for Sub-VT[1]

M7 to M10 to remove the problem of Read SNM by buffering thestored data during a read access.

Thus, the worst-case SNM for this bitcell is the Hold SNMrelated to M1 to M6, which is the same as the 6T Hold SNM for same sizedM1 to M6

M9

M8

M10

M7

11

10T Bitcell Reduces Bitline Leakage

QB=1

RBL=10

QBB heldnear 1 so the leakage current through M8 is reduced

QB=0

RBL=10QBB =1

leakagereducedby stack

Q QB

for iso-VDD, the 10T cell without M10 (a 9T cell) has 50% higher leakage current than the 6T, but adding M10 drops the overhead

to 16%.

[1]

12

Leakage Power Savings with 10T Bitcell

6T memories in 65nm usually at 0.9V or greater (lowest reported is 0.7V)

10T bitcell allows scaling to lower voltages Lower voltage operation reduces leakage power dramatically for unaccessed cells

[1]

13

Bitline Leakage Limits Integration Level

16 bitcells on bitline is best can hope for standard 6T

“1”

“0”

“0”

Bit

-li

ne

[1]

14

Cont’d

BL leakage limits the number of cells on a BL. The 10T bitcell can sustain 256 cells/BL at 0.3V compared to 16 without M10 (6T or 9T). higher level of integration allowed by the 10T cell reduces the peripheral circuits and slightly mitigates the bitcell area overhead[1].

15

10T Bitcell Allows Sub-VT Write

To achieve write in sub-threshold, the virtual supply (VVDD) to the selected cells floats during the write operation

Folded WL shares VVDD

WL

BLBBL

Q QB

VVDD

RBL

RWL

MC

MC MC

MC

VDDon

[1]

16

Cont’d

floating

feedbackrestores ‘1’ to VDD

Floating VDD weakens feedback and allows Write.

A virtual supply voltage (VVDD) that floats during write allows robust write operation into sub-VT (mono-stable butterfly curve). VVDD stops floating while WL_WR remains asserted to restore the ‘1’value to full VDD[1].

17

Test Chip Architecture

• 256 rows and 128 columns per block

• Static CMOS peripherals

• Separate WL VDD for boosting

• Assumed 1x1 redundancy

• Simulation: Operates at 300mV across all process corners from 0 to 100oC

X8

MC

writeBK

prechBK

VDDfloatEn MP<r>

I<c>

RD

column<c>

writeBK

WL_RD

writeB

K

EN

DIO<c>

MC

row<r>

BLB

BL

WL_WR1

MC

128MCMC

VVDD<r>

BL_RD

256

WLglobal BKsel

Address<0:10>

8:256 3:8

BKsel

WLglobal

1[1]

18

256Kb 65nm Sub-VT memory

Test chip addressing the sub-VT problems using 10T bitcell:1.89mm by 1.12mm.

Chip functions to below 400mV, holds without error to <250mV: At 400mV, 3.28mW and 475kHz at 27oC.

Reads without error to 320mV (27oC) and 360mV (85oC).Write without error to 380mV (27oC) and 350mV (85oC).

[1]

19

Simulation results

Chip functioned correctly to below 400mV. Scope plot shows 300mVoperation; at this low voltage, some bit errors were observed[1].

20

Power Measurements

Relative to 0.6V 6T SRAM, 2.2X less leakage power at 0.4V and 3.3X less leakage power at 0.3V >60X less leakage power than 1.2V

[1]

21

Active Energy Savings with 10T Bitcell

200MHz at 1.2V

6T memories in 65nm usually at 0.9V or greater (lowest reported is 0.7V).

Operating 10T bitcell at lower voltages saves energy.

10T memory can provide high frequency operation at higher voltages when necessary.

[1]

22

VDD Scaling Limits

Redundancy and/or boosted WL account for mismatch

1x1 redundancy and WL boosting:

Read works to 320mVWrite works to

380mV

1 column (of 1024)

2 cols

3 cols

Read Bit Errors

Write Bit Errors

1 row (of 2048)

4 rows

5 rows

[1]

23

Conclusions

Standard 6T approach limited to ~0.6-0.7V and 16 cells per bitline.

Proposed 10T bitcell shows sub-threshold operation with overall power and energy savings.

Sub-VT memory requires circuits and architectures to manage variability and low Ion/Ioff.

24

A read SNM free SRAM

• decreases in Read SNM in conventional SRAM cells:

• When SNM>0mV, stable data retention is still achieved even

though the voltage at Node V1 may slightly exceed “0”.

• When SNM<0mV, however, reversal data is overwritten.

1) Node V1 voltage greatly exceeds “0”.

2) Node V2 voltage falls below “1” because Node V1 voltage reaches

the CMOS inverter logic threshold voltage (P2, N2).

3) The fall in Node V2 voltage raises Node V1 voltage further

resulting in the overwriting of reversal data. [3]

25

Cont’d

[3]

26

Proposed read SNM free SRAM

N5 is added between Node V2 and NMOS transistor N2.

When the cell is not accessed /WL is high and when the cell is accessed /wl is low.

N5 prevents V2 from decreasing and thus the data bit is not reversed even if SNM equals zero.

Period of /activation is less than V2 retention time.

[3]

27

Cont’d

A useless gap equal to one transistor results since 7 is a prime number.

A PMOS and an NMOS transistor are placed, respectively, in GAP (P) and GAP (N), between two L-shaped SRAM cells.

Solution: combining the SRAM cell and the sensing circuit.

[3]

28

Measurement results

Organization:Clock access time:

Power consumption:Supply voltage:Process technology:

Macro size:Cell size:

4Kword x 16b1.2 ns, at 1.0 V20 ns, at 0.5 V12.9 mA/GHz, at 1.0 V1.0V to 0.44 V90-nm ASPLA CMOS,NMOS Vth: 0.32V,PMOS Vth: -0.33V0.4 mm x 0.7 mm2.09 μm2 (based on logic rules)

Vdd-min decreases with increasing temperature. Vdd-min is determined by the write SNM, which, unlike Read SNM, improves with decreasing Vth levels in NMOS transistors, and Vth decreases with increasing temperature.

[3]

29

SRAM macro layout and chip microphotograph

Since both Write SNM and SRAM cell current improve with decreasing Vth levels in NMOS transistors, it is possible to achieve even higher-speed and lower- Vdd operations by reducing Vth levels below 0.32V,

[3]

30

REFERENCES

[1]. A 256kb Sub-threshold SRAM in 65nm CMOS

Benton H. Calhoun, Anantha Chandrakasan Massachusetts Institute of Technology, Cambridge, MA ISSCC 2006 / SESSION 34 / SRAM / 34.4

[2]. Analyzing Static Noise Margin for Subthreshold

SRAM in 65nm CMOS Benton H. Calhoun and Anantha Chandrakasan MIT, 50 Vassar St 38-107, Cambridge, MA, 02139 USA {bcalhoun,anantha}@mtl.mit.edu

[3]. A Read-Static-Noise-Margin-Free SRAM

Cell for Low-Vdd and High-Speed Applications

ISCC 2005 / SESSION 26 / STATIC MEMORY / 26.3Koichi Takeda1, Yasuhiko Hagihara1, Yoshiharu Aimoto2, Masahiro Nomura1, Yoetsu Nakazawa1, Toshio Ishii2, Hiroyuki Kobatake

31

THANK YOU…