Professional Documents
Culture Documents
Abstract — We experimentally demonstrate the first the area of a FET), demonstrating the potential to realize
static random-access memory (SRAM) arrays based on a sub-3-nm node technology [6]. Furthermore, the CNFETs
carbon nanotube (CNT) field-effect transistors (CNFETs). are a rapidly maturing nanotechnology, as complete digital
We demonstrate 1 kbit (1024) 6 transistor (6T) SRAM arrays
fabricated with complementary metal-oxide-semiconductor systems [7]–[13], and complex analog and mixed-signal
(CMOS) CNFETs (totaling 6144 p- and n-type CNFETs), with circuits [14], [15] have been experimentally demonstrated.
all 1024 cells functioning correctly without any per-unit Despite these increasingly complex CNFET demonstra-
customization. Moreover, we show the first demonstration tions, a critical component of digital systems—static random
of CNFET CMOS 10T SRAM cells, capable of operating at access memory (SRAM) memory arrays—had never been
highly scaled voltages down to 300 mV. We characterize
the CNFET CMOS SRAM and demonstrate robust opera- demonstrated. Prior work either has realized only individ-
tion by writing and reading multiple patterns (to both the ual, isolated CNT-based SRAM cells [16] or has relied on
kbit arrays as well as the 10T SRAM cells), measuring processing that is not compatible with silicon CMOS (e.g.,
SRAM variations in read, write, and hold margins and relying on air-reactive, ionic, non-solid-state CNT doping
repeat cycling of cells. Moreover, due to the low-temperature processes [17]–[24]). Here, we highlight the first demonstra-
back-end-of-line (BEOL)-compatible CNT-specific process-
ing, CNFET SRAM enables new opportunities for digital tion of the kbit 6T CNFET CMOS SRAM arrays. We demon-
systems, since: 1) CNFET SRAM can be fabricated directly strate all SRAM cells within the array functioning correctly
on top of computing logic to realize three-dimensional without relying on any customization, calibration, or correc-
integrated circuits; and 2) CNFET circuits can utilize metal tion of any sort. This is the largest reported CNFET CMOS cir-
routing both above and below the CNFET device layer (e.g., cuit demonstrated to-date, containing 6144 CMOS CNFETs.
as in our demonstration which utilizes buried power rails,
whereby the power rails are fabricated underneath the FETs To further demonstrate the maturity of CNFET CMOS, we
while metal routing is fabricated above the FETs), providing also demonstrate the first 10T CNFET CMOS SRAM cells,
opportunities for further SRAM density scaling. which can operate at highly-scaled supply voltages down to
Index Terms — Carbon nanotubes (CNTs), carbon 300 mV. Importantly, all design and processing is 1) wafer-
nanotube field-effect transistors (CNFETs), digital logic, scale (performed on 150-mm substrates), 2) VLSI-compatible
static random-access memory (SRAM). (no per-unit customization), and 3) silicon-CMOS-compatible
(leveraging only solid-state and silicon-compatible materials
I. I NTRODUCTION and conventional fabrication process steps).
This work is an extension of [25]. In addition to summa-
C ARBON nanotube (CNT) field-effect transistors
(CNFETs) are a promising emerging nanotechnology
for next-generation energy-efficient digital very-large-scale-
rizing those results, we include additional 6T SRAM charac-
terization, as well as the first experimental demonstrations of
CNFET CMOS 10T SRAM cells.
integration (VLSI) circuits. Owing to their simultaneously
ideal electrostatic control (due to the ultrathin ∼1-nm
diameter of a CNT) and high carrier transport [1], [2], it is II. CNFET CMOS SRAM FABRICATION
projected that the CNFETs can improve the energy-delay
The CNFET CMOS fabrication process flow is illustrated
product (EDP, a metric of energy efficiency) by an order
in Fig. 1. The starting substrate is a 150-mm silicon wafer
of magnitude versus silicon complementary metal-oxide-
with ∼1 µm thermal SiO2 . Prior to any CNFET fabrication,
semiconductor (CMOS) [3]–[5]. In addition to these EDP
we fabricate a bottom metal layer of metal as buried power
benefits, CNFET digital logic has been fabricated at a record
rails. An inter-layer dielectric (ILD) is deposited followed
scaled 30 nm contacted gate-pitch (CGP, a key metric defining
by defining metal vias down to the buried power rails. The
Manuscript received August 26, 2019; accepted September 26, next metal layer is defined, which acts as both metal routing
2019. Date of publication November 13, 2019; date of current ver- (for the wordlines) as well as the bottom metal gates for the
sion November 27, 2019. This work was supported in part by Analog
Devices, Inc., DARPA 3DSoC and in part by the Air Force Research CNFETs. To fabricate the CNFETs, a high-k gate dielectric
Laboratory. The review of this article was arranged by Editor K. Uchida. (HfO2 ) is deposited through atomic layer deposition (ALD) at
(Corresponding author: Pritpal S. Kanhaiya.) 200 ◦ C, followed by CNT deposition. We deposit the CNTs
The authors are with the Massachusetts Institute of Technology,
Cambridge, MA 02139 USA (e-mail: pkanhaiy@mit.edu). uniformly across the 150 mm substrate through a solution-
Digital Object Identifier 10.1109/TED.2019.2945533 based deposition, whereby the wafer is submerged in a solvent
0018-9383 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on December 15,2021 at 16:53:13 UTC from IEEE Xplore. Restrictions apply.
5376 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 66, NO. 12, DECEMBER 2019
The next metal layer (Pt) is defined, which acts as both the
CNFET source and drain contacts for the p-type CNFETs
and additional metal routing (bit/bit_b lines). The finished p-
type CNFETs are protected and passivated with 100 nm SiO2 .
To fabricate the n-type CNFETs, we deposit the next metal
layer (Ti) as the n-type CNFET source and drain contacts,
followed by ALD deposition of HfO2 (which electrostatically
dopes the CNTs to achieve the correct threshold voltage,
see [28] for details). The HfO2 electrostatic doping film is
etched off of the p-type CNFETs (which are protected by
the SiO2 passivation), leaving our final CNFET CMOS. The
full process contains three metal layers and uses only silicon-
CMOS compatible materials. Key aspects of this fabrication
process are:
1) The buried metal layers beneath the CNFET device layer
can be fabricated using conventional back-end-of-line
(BEOL) metallization processes.
2) Such buried metal layers are very challenging to real-
ize with conventional silicon CMOS due to high-
temperature processing required for fabricating silicon
CMOS (>1000 ◦ C for steps such as dopant activation
annealing [29]).
3) In contrast, the low-temperature (<400 ◦ C) CNFET
processing naturally enables such buried metal layers
and new three-dimensional circuit architectures. Specif-
ically, the room-temperature solution CNT deposition
decouples the high-temperature CNT synthesis (typically
>800 ◦ C) from the wafer fabrication, and the CNFET
CMOS doping process (leveraging both metal source
and drain work function engineering as well as high-
k electrostatic doping) never exceeds 250 ◦ C.
4) While our experimental demonstration of the CNFET
CMOS SRAM only leverages a single buried metal
layer (utilized as buried power rails), the same BEOL
metallization process can be repeated to realize multiple
buried metal layers (in addition to multiple conventional
BEOL metal layers above the device layer).
5) The entire fabrication process is wafer-scale, solid-state,
and BEOL compatible (<400 ◦ C). Thus, while we fabri-
Fig. 1. Process flow for the CNFET CMOS SRAM. (a) M1, the first metal cate our CNFET CMOS SRAM over silicon substrates,
layer, which consists of the VDD and VSS lines, is defined (5 nm Ti/45 nm the substrate could be silicon CMOS circuitry, additional
Pt) via physical vapor deposition (PVD). (b) Oxide spacer deposition layers of CNFET CMOS circuitry, etc.
followed by etch and metal fill for inter-layer via (ILV) definition. (c) M2,
the second metal layer, is defined (5 nm Ti/70 nm Pt) for the wordlines Fig. 2 shows the circuit schematic as well as a three-
and back gates of the CNFETs. (d) High-k gate dielectric deposition dimensional schematic of a single 6T CNFET CMOS SRAM
(30 nm HfO2 ) via ALD, and additional ILV definition. (e) CNTs solution- cell. Fig. 3 shows the scanning electron microscopy (SEM)
deposited at room temperature uniformly across the wafer followed
by oxygen plasma to etch away CNTs outside the active area (e.g., images of a fabricated 1 kbit CNFET CMOS 6T SRAM array
channel). (f) M3, the third metal layer, consists of bit, bit_b and PMOS as well as a single 6T SRAM. Figs. 2 and 3 highlight the
source/drain definition (2 nm Ti/98 nm Pt) as well as NMOS source/drain buried metal process whereby the power lines (VDD and VSS )
definition (100 nm Ti). (g) Passivation over the PMOS (100 nm SiO2 ).
(h) Electrostatic doping over the NMOS (20 nm HfO2 ). are routed directly beneath the CNFET CMOS SRAM array,
with the CNFETs and signal routing fabricated directly verti-
cally overlapping (on the same substrate without any die- or
wafer-bonding). Such three-dimensional circuit architectures
containing dispersed CNTs (the CNTs are pre-purified within can provide additional future opportunities for further SRAM
this solution to achieve >99.99% semiconducting CNTs, density scaling and reduced routing congestion. The SRAM
see [26], [27] for details). Following CNT deposition, the density in this demonstration is limited due to the fact that
active area (e.g., channel area) of the CNFETs is defined the CNFETs are fabricated at a relaxed ∼1-µm technology;
through traditional photolithography and all CNTs outside this is only due to the limitations of academic fabrication
of the active regions are etched away using oxygen plasma. facilities, as this same CNFET fabrication process has been
Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on December 15,2021 at 16:53:13 UTC from IEEE Xplore. Restrictions apply.
KANHAIYA et al.: CNT-BASED CMOS SRAM: 1-KBIT 6T SRAM ARRAYS AND 10T SRAM CELLS 5377
Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on December 15,2021 at 16:53:13 UTC from IEEE Xplore. Restrictions apply.
5378 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 66, NO. 12, DECEMBER 2019
Fig. 7. False colored SEM image of the (a) fabricated 10T CNFET CMOS
SRAM cell and (b) 10T SRAM cell schematic.
Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on December 15,2021 at 16:53:13 UTC from IEEE Xplore. Restrictions apply.
KANHAIYA et al.: CNT-BASED CMOS SRAM: 1-KBIT 6T SRAM ARRAYS AND 10T SRAM CELLS 5379
V. C ONCLUSION
This work realizes the first CNFET CMOS SRAM arrays
(1 kbit 6T SRAM) comprising 6144 CMOS CNFETs, as well
as the first 10T CNFET CMOS SRAM cells operating at
highly scaled supply voltages down to 300 mV. We also
demonstrate buried power rails for further SRAM density
scaling opportunities. Taken together, this work is a major step
Fig. 8. 10T cell-level characterization. (a) Write margin (∼0.35 V) and toward demonstrating the feasibility of a future CNT-based
(b) read margin (∼0.23 V) from a typical CNFET CMOS 10T SRAM technology, realizing a critical component of future energy-
cell (VDD = 1 V). Time-dependent waveforms at (c) 1 V VDD and (d) efficient CNT digital systems.
300 mV VDD . There is still positive read and write margins at VDD = 300
mV, a requirement for digital logic (∼0.13 and ∼0.02 V, respectively).
Below 300 mV, noise margin violations result in incorrect operation. Clock ACKNOWLEDGEMENT
frequency is limited by the off-chip data acquisition setup that records
the output waveform. The measurements in (c) and (d) show first writing This article is an extension of the VLSI 2019 conference
a “1” followed by four nondestructive reads, followed by writing a “0”
followed by four nondestructive reads. The noise in the waveform is
article: P. S. Kanhaiya et al., “1 kbit 6T SRAM Arrays in
from measuring the cell while the bit lines are floating, illustrating that Carbon Nanotube FET CMOS,” IEEE VLSI 2019.
the access transistors [CNFET “A1” and “A2” in Fig. 7(b)] provide ideal
electrical isolation of the cell from the bitline.
R EFERENCES
margin low (RML/HML) are defined in (4)–(6) as [1] G. J. Brady, A. J. Way, N. S. Safron, H. T. Evensen, P. Gopalan,
and M. S. Arnold1, “Quasi-ballistic carbon nanotube array transistors
(dr) (ld)
RMH or HMH = VOH − VIH (4) with current density exceeding Si and GaAs,” Sci. Adv., vol. 2, no. 9,
Sep. 2016, Art. no. e1601240, doi: 10.1126/sciadv.1601240.
(ld) (dr)
RML or HML = VIL− VOL (5) [2] A. Javey, J. Guo, Q. Wang, M. Lundstrom, and H. Dai, “Ballistic
(dr) (ld) (ld) (dr) carbon nanotube field-effect transistors,” Nature, vol. 424, no. 6949,
RM or HM = min(VOH − VIH ¸VIL − VOL ). (6) pp. 654–657, Aug. 2003, doi: 10.1038/nature01797.
Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on December 15,2021 at 16:53:13 UTC from IEEE Xplore. Restrictions apply.
5380 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 66, NO. 12, DECEMBER 2019
[3] G. Hills et al., “Understanding energy efficiency benefits of [18] S. Y. Lee et al., “Scalable complementary logic gates with chemically
carbon nanotube field-effect transistors for digital VLSI,” IEEE doped semiconducting carbon nanotube transistors,” Acs Nano, vol. 5,
Trans. Nanotechnol., vol. 17, no. 6, pp. 1259–1269, Nov. 2018, no. 3, pp. 2369–2375, Mar. 2011, doi: 10.1021/nn200270e.
doi: 10.1109/TNANO.2018.2871841. [19] Z. Zhang et al., “Doping-free fabrication of carbon nanotube based
[4] M. M. S. Aly et al., “Energy-efficient abundant-data computing: The ballistic CMOS devices and circuits,” Nano Lett., vol. 7, no. 12,
N3XT 1,000x,” Computer, vol. 48, no. 12, pp. 24–33, Dec. 2015, pp. 3603–3607, Nov. 2007, doi: 10.1021/nl0717107.
doi: 10.1109/MC.2015.376. [20] D. Shahrjerdi, A. D. Franklin, S. Oida, J. A. Ott, G. S. Tulevski,
[5] L. Wei, D. J. Frank, L. Chang, and H.-S. P. Wong, “A non- and W. Haensch, “High-performance air-stable n-type carbon nan-
iterative compact model for carbon nanotube FETs incorporating otube transistors with erbium contacts,” ACS Nano, vol. 7, no. 9,
source exhaustion effects,” in IEDM Tech. Dig., Dec. 2009, pp. 1–4, pp. 8303–8308, Sep. 2013, doi: 10.1021/nn403935v.
doi: 10.1109/IEDM.2009.5424281. [21] L. Ding et al., “Y-contacted high-performance n-type single-walled
[6] T. Srimani, G. Hills, M. D. Bishop, and M. M. Shulaker, “30-nm carbon nanotube field-effect transistors: Scaling and comparison with Sc-
Contacted Gate Pitch Back-Gate Carbon Nanotube FETs for Sub-3- contacted devices,” Nano Lett., vol. 9, no. 12, pp. 4209–4214, Nov. 2009,
nm Nodes,” IEEE Trans. Nanotechnol., vol. 18, pp. 132–138, 2019, doi: 10.1021/nl9024243.
doi: 10.1109/TNANO.2018.2888640. [22] J.-L. Xu et al., “Efficient and reversible electron doping of
[7] M. M. Shulaker et al., “Carbon nanotube computer,” Nature, vol. 501, semiconductor-enriched single-walled carbon nanotubes by using
no. 7468, pp. 526–530, Sep. 2013, doi: 10.1038/nature12502. decamethylcobaltocene,” Sci. Rep., vol. 7, no. 1, Jul. 2017, Art. no. 6751,
[8] M. M. Shulaker et al., “Monolithic 3D integration of logic doi: 10.1038/s41598-017-05967-w.
and memory: Carbon nanotube FETs, resistive RAM, and sili- [23] M.-L. Geier, K.-K. Moudgil, S. Barlow, S.-R. Marder, and
con FETs,” in IEDM Tech. Dig., Dec. 2014, pp. 27.4.1–27.4.4, M.-C. Hersam, “Controlled n-type doping of carbon nanotube transistors
doi: 10.1109/IEDM.2014.7047120. by an organorhodium dimer,” Nano Lett., vol. 16, no. 7, pp. 4329–4334,
[9] M. M. Shulaker et al., “Carbon nanotube circuit integration up to sub-20 Jun. 2013, doi: 10.1021/acs.nanolett.6b01393.
nm channel lengths,” ACS Nano, vol. 8, no. 4, pp. 3434–3443, Mar. 2014, [24] M.-G. Zhu, Z. Zhang, and L.-M. Peng, “High-performance and
doi: 10.1021/nn406301r. radiation-hard carbon nanotube complementary static random-
[10] M. M. Shulaker et al., “Sensor-to-digital interface built entirely with access memory,” Adv. Electron. Mater., vol. 5, no. 7, Jul. 2019,
carbon nanotube FETs,” IEEE J. Solid-State Circuits, vol. 49, no. 1, Art. no. 1900313, doi: 10.1002/aelm.201900313.
pp. 190–201, Jan. 2014, doi: 10.1109/JSSC.2013.2282092. [25] P. S. Kanhaiya, C. Lau, G. Hills, M. Bishop, and M. M. Shulaker, “1 Kbit
[11] H. Wei, M. Shulaker, H.-S. P. Wong, and S. Mitra, “Monolithic 6T SRAM arrays in carbon nanotube FET CMOS,” in Proc. Symp. VLSI
three-dimensional integration of carbon nanotube FET complemen- Technol., Jun. 2019, pp. T54–T55, doi: 10.23919/VLSIT.2019.8776563.
tary logic circuits,” in IEDM Tech. Dig., Dec. 2013, pp. 19–7, [26] H. Wang and Z. Bao, “Conjugated polymer sorting of semiconducting
doi: 10.1109/IEDM.2013.6724663. carbon nanotubes and their electronic applications,” Nano Today, vol. 10,
[12] L. Ding et al., “CMOS-based carbon nanotube pass-transistor logic no. 6, pp. 737–758, Dec. 2015, doi: 10.1016/j.nantod.2015.11.008.
integrated circuits,” Nature Commun., vol. 3, Feb. 2012, Art. no. 677. [27] J. W. T. Seo et al., “Diameter refinement of semiconducting arc discharge
[13] M. Shulaker et al., “Experimental demonstration of a fully digi- single-walled carbon nanotubes via density gradient ultracentrifugation,”
tal capacitive sensor interface built entirely using carbon-nanotube J. Phys. Chem. Lett., vol. 4, no. 17, pp. 2805–2810, Aug. 2013, doi:
FETs,” in IEEE ISSCC Dig. Tech. Papers, Feb. 2013, pp. 112–113, 10.1021/jz4013596.
doi: 10.1109/ISSCC.2013.6487660. [28] C. Lau, T. Srimani, M. D. Bishop, G. Hills, and M . M. Shu-
[14] A. G. Amer, R. Ho, G. Hills, A. P. Chandrakasan, and M. M. Shulakerm, laker, “Tunable n-type doping of carbon nanotubes through engineered
“SHARC: Self-healing analog with RRAM and CNFETs,” in atomic layer deposition HfOx films,” ACS Nano., vol. 12, no. 11,
IEEE ISSCC Dig. Tech. Papers, Mar. 2019, pp. 470–472, pp. 10924–10931, Oct. 2018, doi: 10.1021/acsnano.8b04208.
doi: 10.1109/ISSCC.2019.8662377. [29] P. Batude et al., “Advances, challenges and opportunities in 3D CMOS
[15] R. Ho, C. Lau, G. Hills, and M. Shulaker, “Carbon nanotube CMOS sequential integration,” in IEDM Tech. Dig., Dec. 2011, pp. 151–154,
analog circuitry,” IEEE Trans. Nanotechnol., vol. 18, pp. 845–848, 2019, doi: 10.1109/IEDM.2011.6131506.
doi: 10.1109/TNANO.2019.2902739. [30] D. K. Pradhan, J. Singh, and S. P. Mohanty, Robust SRAM Designs and
[16] M. L. Geier et al., “Solution-processed carbon nanotube thin-film Analysis. New York, NY, USA: Springer, 2012, pp. 33–38.
complementary static random access memory,” Nature Nanotechnol., [31] N. Rahman, G. Dhiman, and B. P. Singh, “Static-noise-margin analysis
vol. 10, no. 11, pp. 944–948, Sep. 2015, doi: 10.1038/nnano.2015. of modified 6t SRAM cell during read operation,” Int. J. Recent Trends
197. Eng. Technol., vol. 8, no. 2, p. 47, Jan. 2013.
[17] H. Wang et al., “Tuning the threshold voltage of carbon nanotube transis- [32] B. H. Calhoun and A. Chandrakasan, “A 256-kb 65-nm
tors by n-type molecular doping for robust and flexible complementary sub-threshold SRAM design for ultra-low-voltage operation,”
circuits,” Proc. Nat. Acad. Sci. USA, vol. 111, no. 13, pp. 4776–4781, IEEE J. Solid-State Circuits, vol. 42, no. 3, pp. 680–688, Mar. 2007,
Apr. 2014, doi: 10.1073/pnas.1320045111. doi: 10.1109/JSSC.2006.891726.
Authorized licensed use limited to: MANIPAL INSTITUTE OF TECHNOLOGY. Downloaded on December 15,2021 at 16:53:13 UTC from IEEE Xplore. Restrictions apply.