You are on page 1of 3

Graphene-Based Plasma Wave Interconnects for On-Chip

Communication in the Terahertz Band


Shaloo Rakheja and Kexin Li
Electrical and Computer Engineering, New York University, NY, USA, E-mail: shaloo.rakheja@nyu.edu

As the communication complexity is exacerbated by dimensional scaling of on-chip components, it becomes important
to investigate communication and transduction mechanisms that can deliver enhanced connectivity, while minimizing
energy dissipated in communication. The communication bottleneck can be mitigated by using graphene-based
plasmonic interconnects in high-performance systems. In this work, we use the phenomenon of Dyakonov-Shur (DS)
instability in a quasi-ballistic GaN high-electron mobility transistor (HEMT) to electrically excite surface plasmon
polaritons (SPPs) in graphene serving as the gate electrode of the HEMT. Information encoded in SPPs is guided
along the graphene waveguide for low-energy on-chip data communication. We evaluate and compare the
performance of plasmonic interconnects against electrical interconnects at scaled technology nodes.
I. Graphene-GaN HEMTs for plasma wave transceivers and waveguides: The design of the plasmonic generator
and detector based on the DS instability [1-2] is shown in Fig. 1(a,b). The 2DEG formed in the GaN HEMT channel
is described using the hydrodynamic equation of motion with reduced external friction. Mathematical details can be
found elsewhere [2]. The device functions as a THz transmitter for constant current boundary condition at the drain
terminal and constant voltage boundary condition at the gate and source terminals. On the other hand, when the HEMT
is biased with a constant gate-source voltage and subject to an incident electromagnetic radiation, it develops a
constant drain-source voltage, thereby acting as a THz detector. The plasma waves in the 2DEG have been
experimentally demonstrated in various transistor structures [3-4]. The DS instability in the HEMT channel is
accompanied by a variation of the dipole moment of charges which create mirror image charges or SPPs in the
multilayer (ML) graphene electrode. The SPPs, polarized in the transverse magnetic direction in the THz band, can
be guided to another plasmonic transistor for energy-efficient local data communication up to 10’s of microns. The
SPP dispersion relationship, !O .;:  at the DS resonant frequency, .;: , in graphene is given as
/1 /2 + .;: (1a)
\ _ 
.;: 2 . 2 .;: /0
!O2 ] /1 !O2 ] /2 ;:
 
%G2 ] %02 )%G (1b)
.;: _ ) ` b%  %0 c
>?? %G >?? G
where /1 /2 are the relative permittivities of the dielectric media surrounding graphene, + . is the dynamical
conductivity of ML graphene, c is the speed of light, vp (v0) is the plasma wave velocity (drift velocity) in the HEMT
channel, the integer N denotes the mode of the plasma wave, and Leff is the HEMT channel length. The plasma wave
> R EQ
velocity %G _   where e is the elementary charge, n0 is the 2DEG concentration that is modulated via the
DVWW 7S
application of gate voltage, meff = 0.2m0 is the effective mass of electrons in GaN, and CG is the gate capacitance. In
Fig. 1c, the resonant frequency is plotted as a function of the ratio #0 9 for Al0.3Ga0.7N/GaN HEMT (Leff = 40 nm).
The ratio #0 9 is modulated via gate voltage as shown in the inset plot of Fig. 1c (simulation details can be found in
[5].) We choose two values of .; = 22 THz and 12.5 THz denoted as points “a” and “b” in Fig. 1c. We only focus on
 =1 mode under the condition that %G  %0 ; this condition is easily satisfied at the THz operation frequency [2]. The
dynamical conductivity of ML graphene depends on the scattering rate, number of layers in the stack, electrostatic
screening length, and the Fermi level in each layer. Using the Drude formalism, the intra-band dynamical conductivity
of ML graphene in the presence of electrostatic screening is given as [6]
:YU[VZ (2a)
&N ] .
+ . _ -N 
. 2 \ &N2
N51
 2 ?N ?N (2b)
-N _ \  \  ] 
!6  !6  !6 
0 (] (2c)
?N _ ?0  ] 
'I

Authorized licensed use limited to: SASTRA. Downloaded on March 20,2024 at 04:30:14 UTC from IEEE Xplore. Restrictions apply.
where &N and ?N are the electron scattering rate and Fermi level, respectively, in the #JA layer of the graphene stack,
?0 is a fitting parameter obtained from experimental calibration, 0 _ nm (vertical spacing between the layers),
and 'I is the electrostatic screening length. The dynamical conductivity is plotted in Fig. 2a as a function of frequency
of operation assuming that the scattering rate is limited by charged carriers and acoustic phonons as shown in Fig. 2b.
An increase in 'I increases both the real and imaginary parts of the graphene conductivity, which offers superior
plasmon propagation characteristics. Combining (1) and (2), the plasmon propagation constant versus frequency is
plotted in Fig. 2(c). A large value of the real part of !O is advantageous for confining plasmons to the graphene surface;
however, a large value of the imaginary part of !O implies higher ohmic losses and reduced propagation length of the
plasmonic signal. Therefore, there exists a trade-off between propagation length and localization of SPPs – an optimal
design of plasmonic waveguide ensures that the ratio of propagation and localization lengths is maximized.
II. Bandwidth, bandwidth density and energy considerations: For local data communication in the plasmonic
domain, we consider the propagation of a narrow-band Gaussian signal centered at the THz resonant frequency, .; .
The resonant frequency for all simulations is chosen as .; _22 THz and 12.5 THz corresponding to Vgs=-1.9V and
Vgs=-2.2 V, respectively, for AlGaN thickness of 12 nm (see inset of Fig. 1c). As shown in Fig. 3a, the envelope of
the Gaussian signal suffers from a decrease in amplitude and pulse width broadening as it propagates along the
plasmonic waveguide. The reduction in amplitude is a direct result of ohmic losses in graphene, which implies a higher
imaginary part of propagation constant. The pulse width broadening increases monotonically with an increase in the
propagation distance as shown in Fig. 3b. Due to ohmic losses in the media, the group delay acquires a diffusive
character. That is, the group delay of the Gaussian signal increases quadratically with waveguide length at longer
length scales as shown in Fig. 3c. If + . were purely imaginary (negligible &N in (2a)), issues related to propagation
losses can be mitigated. However, in practical graphene structures, &N is indeed finite [6].
Due to pulse width broadening, successive pulses launched into the waveguide may overlap if the inter-pulse timing
1 =PT 1
is not sufficient. The maximum achievable bitrate or bandwidth is given as < a _  , where $@ is the
MJX =JX MPT
group delay and is a nonlinear function of frequency. In Fig. 4a, the bandwidth of plasmonic and electrical
interconnects is compared as a function of interconnect length for two values of the effective resistivity of copper and
assuming that electrical interconnects operate in the RC regime. In general, the bandwidth of both electrical and
plasmonic interconnects degrades with an increase in interconnect length; however, this degradation is much higher
for electrical interconnects due to their diffusion dominated transport. A straightforward way to improve the bandwidth
is to increase the width of the wire; this will result in a concomitant increase in bandwidth density (BWD = Fb/(2W))
only when the aspect ratio of the wire is fixed [7]. In Fig. 4b, we show the impact of interconnect width on BWD of
plasmonic and electrical interconnects computed at a fixed location on the wire. For resistance-dominated electrical
interconnects, BWD increases with an increase in width for aspect ratio chosen as 2. In the plasmonic domain, the wire
width is essentially determined by the resonant frequency to avoid edge plasmon modes and leakage of the plasmonic
signal into the surrounding dielectric media. Therefore, we choose the width of plasmonic waveguide as  _
b!O c, where !O is obtained from (1). As such, the ratio <  is independent of width once the frequency of
operation is fixed. Finally, we also compare the energy-per-bit (EPB) of electrical and plasmonic interconnects and
the results are plotted in Fig. 4c. Within the shot-noise limited transmission in the plasmonic domain, the EPB is given
1
as  _ .; " bBEJ GHFG c, where " denotes the average number of plasmons needed to encode one bit
2
of information, and Lint (Lprop) is the interconnect (propagation) length. For a specified bit-error rate (BER) in signal
detection, " _ ]b ^c. In this work, we choose  _ 430 , which yields " `68. For electrical
1 2
interconnects, EPB is given as >C _ JL \ HL \ K BEJ 88 , where JL HL is the transmitter (receiver)
2
capacitance, K is the per-unit-length wire capacitance, and 88 is the supply voltage. Unlike electrical interconnects
for which EPB increases only linearly with interconnect length, the EPB of plasmonic interconnects shows an
exponential dependence on interconnect length. As shown in Fig. 4c, plasmonic interconnects outperform their
electrical counterparts in terms of EPB at short length scales up to few 10’s of micrometers when the electrostatic
screening length, 'I is finite. The effect of 'I is quite significant on EPB; therefore, a graphene stack where the
electrical coupling between the parallel layers is strong such that 'I is large is preferred for waveguiding applications.
III. Conclusion: The paper quantifies the opportunities of graphene plasmonics for on-chip communication in the
THz band. We focus on the electrical generation of plasma waves in a quasi-ballistic GaN HEMT governed by the
Dyakonov-Shur instability. We show that plasmonic interconnects offer superior bandwidth, bandwidth density, and
energy-per-bit when compared against electrical interconnects at scaled technology nodes.
Acknowledgment: This work was supported in part by the NSF (Award # CCF-1565656) and by The Boeing Company.

Authorized licensed use limited to: SASTRA. Downloaded on March 20,2024 at 04:30:14 UTC from IEEE Xplore. Restrictions apply.
References: [4] F. Teppe et al., APL 87, p. 52107, 2005.
[1] M.I. Dyakonov and M.S. Shur, IEEE T-ED, vol.43, pp. 1640- [5] X.-D. Wang et al., IEEE T-ED, vol. 59, no. 5, pp. 1393-1401, 2012.
1645, 1996. [6] S. Rakheja, IEEE T-NANO, vol. 15, no. 6, pp. 936-946, 2016.
[2] M.I. Dyakonov and M.S. Shur, APL 87, p. 111501, 2005. [7] D. Miller and H. Ozaktas, Journal of Parallel and Distributed Computing, vol.
[3] W. Knap et al., APL 85, pp. 657-677, 2004. 41, pp. 42-52, 1997.
VG VG
SPP SPP
incoming
Graphene Graphene
         
ID  
D S D S VS
VS 


ΔV
VDS
 
Fig. 1a: THz transmitter by coupling plasma waves Fig. 1b: THz receiver in which incoming radiation
in the 2DEG of HEMT with surface plasmon excites instability in 2DEG and creates a DC drain-
polaritons (SPPs) in graphene. source voltage.

Fig. 1c: Resonant frequency versus n0/CG. Inset
shows n0/CG for two thicknesses of AlGaN.
3 12 3
4 ×10
10 10
Solid: λs → ∞ Solid: λs → ∞
Imaginary Solid:  avg = 9
Dashed: λs = 1 nm Dashed: λs = 1 nm
102 3.5 Dashed:  avg = 1 N = 10
15
m
-2
Nlayer = 5 imp Nlayer = 5
102
E f0 = 0.25 eV 3 E f0 = 0.25 eV
σ(ω)/σuni

1
10
Γ (s )

2.5

0
-1

Real
101

k /k
ρ
10
0 2 Nimp = 10
14
m
-2

Real Imaginary
1.5
-1 100
10 1
-2 0.5 -1
10 0.2 0.4 0.6 0.8 10
0 10 20 30 40 0 10 20 30 40
Ef0 (eV)
Frequency (THz)   Frequency (THz) 
Fig. 2a: Dynamical conductivity of multilayer Fig. 2b: Scattering rate versus Fermi level for Fig. 2c: Real and imaginary part of SPP
graphene with five layers by considering only intra- various values of charged impurity concentration propagation in graphene versus frequency for
band transitions. and dielectric permittivities. various values of screening length, 'I .
1 E = 0.25 eV 14
1 f0
z = 10 μm 0 Nlayer = 5 Dashed: Linear component
z = 100 μm 10 12
0.8 N = 10
15
m
-2
|F(z,t)|

z = 200 μm imp
N =5
layer
0.5 10 E = 0.25 eV
f0
ω = 12.5 THz
[τ/τ -1]

0.6 R λ s = 1 nm
|F(z,t)|

tg (ps)

8
10-2
0

0
0 0.5 1
0.4 Time (s) ×10
-11 6
ω = 22 THz; λ = 1 nm
R s
ω = 22 THz ωR = 12.5 THz; λs = 1 nm
R 4
0.2 z = 10 μm 10
-4 ω = 22 THz; λ → ∞
R s ω = 22 THz
z = 100 μm ω = 12.5 THz; λ → ∞ 2 R
z = 200 μm R s ω
R
= 12.5 THz
0 0 1 2 0
0 0.5 1 10 10 10 0 20 40 60 80 100
Times (s) ×10-11  Length (μm)  Length (μm) 
Fig. 3a: Broadening and amplitude reduction of a Fig. 3b: Normalized pulse width broadening as a Fig. 3c: Group delay of the Gaussian signal as a
Gaussian signal as it propagates along the function of the waveguide length. ,0 (,) is the function of the waveguide length. Diffusive
waveguide at two values of .; . unbroadened (broadened) pulse width. character is noted for 'I = 1 nm.
10
10
Electrical (V = 0.7 V)
DD
8 ω R = 22 THz
10 ω = 12.5 THz
R
Energy/k T
B

6
10

4
10

2
10
N =5
layer
0 E = 0.25 eV
f0
10 0 1 2
10 10 10
  Length (μm) 
Fig. 4a: Bandwidth versus interconnect length for Fig. 4b: Bandwidth density for both electrical and Fig. 4c: Energy-per-bit versus interconnect length
electrical and plasmonic domains. *>?? is the plasmonic interconnects evaluated at 100 µm for electrical and plasmonic interconnects. VDD is
effective resistivity of copper. distance for two values of .;  the supply voltage.

Authorized licensed use limited to: SASTRA. Downloaded on March 20,2024 at 04:30:14 UTC from IEEE Xplore. Restrictions apply.

You might also like