Content uploaded by Zvi Or-Bach
Author content
All content in this area was uploaded by Zvi Or-Bach on Jan 11, 2015
Content may be subject to copyright.
Monolithic 3D-ICs with Single Crystal Silicon Layers
Deepak C. Sekar and Zvi Or-Bach
MonolithIC 3D Inc., San Jose, CA 95124, USA E-mail: deepak@monolithic3d.com (Invited Paper)
Abstract
Approaches to obtain monolithic 3D logic and memory ICs are
proposed in this paper. 3-4x higher memory density with similar
litho cost can be obtained with monolithic 3D memories, while
benefits similar to a generation of scaling can be obtained with
monolithic 3D logic by doubling the number of device layers.
Well-known, manufacturing-friendly materials, process steps
and device structures are used.
Introduction
3D-ICs with monolithically stacked transistor layers provide
ultra-dense through-silicon connections and short wires. In
addition, they often utilize the same litho step to pattern multiple
layers of devices, thereby reducing litho cost. A few trends are
making monolithic 3D-ICs increasingly attractive. Fig. 1 shows
litho tool cost has increased exponentially over the past 40 years.
Next-generation litho is now quite expensive and risky. Fig. 2
indicates wire RC delays are a key bottleneck in today’s chips. A
28nm GPU, for instance, requires several times more energy for
communicating data than for computation. This trend is expected
to get worse with scaling (Fig. 2). Down-scaling requires major
changes to transistors too, with disruptive innovations such as
high k/metal gate and Finfets required every few years.
Ion-Cut Technology
Single crystal silicon for stacked device layers of Monolithic
3D-ICs can be obtained with ion-cut technology, which involves
hydrogen implantation, wafer bonding and cleave (Fig. 3). Ion-
cut has been demonstrated below 400oC [3], and is well-known
due to its two decades of use for SOI wafer manufacturing.
Monolithic 3D Memories
Flash memory, which scales faster than logic and DRAM, may
hit the limits of traditional scaling first. Toshiba, Samsung,
Micron and Hynix have all invested aggressively in polysilicon-
based monolithic 3D flash memories, and have them on their
roadmaps [4][5]. Fig. 4 shows our proposal for single crystal
silicon 3D flash memories with litho steps shared among
multiple device layers. Single crystal silicon flash memory cells
have ~5x higher mobility, better off-characteristics and lower
variability compared with poly cells, all of which help with
multi-bit storage. As shown in Fig. 4, ion-cut is utilized
repeatedly to produce multiple layers of single crystal silicon
atop peripheral circuits with tungsten wiring, following which
shared litho steps are utilized to define NAND flash memory
strings and produce contacts [5]. Fig. 5 shows memory capacity
estimations based on [5]. For a 140mm2 die, the structure in Fig.
4 can provide 256Gbit chips compared with 64Gbit and 128Gbit
for conventional scaled NAND and polysilicon-based 3D
vertical NAND flash respectively [5]. This is due to the cost
advantages of sharing litho steps across multiple memory layers
as well as the multi-bit storage capacity of single crystal silicon
cells. Maximum aspect ratios for etch and deposition in Fig. 4
are 16:1 compared to higher than 50:1 for poly-based 3D vertical
NAND flash, allowing easier manufacturing. Note that ion-cut
could be applied to any 3D flash memory with horizontal
channels, and is not limited to the exemplary architecture [6]
shown in Fig. 4.
Monolithic 3D architectures with shared litho steps have been
developed by the authors for floating-body DRAM and resistive
memories too [7]. Single crystal silicon is an enabler for these
applications. The primary concern with applying ion-cut to
memories is cost. Our analysis reveals that with re-use of
substrates, ion-cut could cost as low as $60 per layer, making it
affordable. Encouragingly, some companies in the cost-sensitive
solar industry are adopting ion-cut today.
Monolithic 3D Logic
Much of today’s 3D logic stacks utilize chips processed
separately with high temperatures that are then thinned, bonded
and connected to each other. Silicon thickness of thinned die and
misalignment during bonding are concerns with this approach.
The ITRS predicts minimum TSV diameter around 1um between
2009 and 2015, indicating difficulties scaling to small TSV sizes
needed for many applications. Monolithic 3D-ICs could be a
solution. Fig. 6 indicates the main barrier to creating high-
quality transistors at Cu/low k compatible temperatures (sub-
400oC) is dopant activation. Fig. 7 describes one approach to
overcome this problem, which utilizes recessed channel
transistors. These have been used in DRAM manufacturing since
the 90nm node, and are known to be competitive with standard
planar transistors [9]. As can be seen in Fig. 7, high temperature
dopant activation steps are conducted before transferring bilayer
n+/p silicon layers atop Cu/low k using ion-cut. The transferred
layers are unpatterned, therefore no misalignment issues occur
while bonding. Following bonding, sub-400oC etch and
deposition steps are used to define the recessed channel
transistor. This is enabled by the unique structure of the device.
These transistor definition steps can use alignment marks of the
bottom Cu/low k stack since transferred silicon films are thin
(usually sub-100nm) and transparent. Sub-50nm through-silicon
connections can be produced due to the excellent alignment.
To investigate the chip-level impact of monolithic 3D-ICs, an
open-source 2D/3D chip simulator called IntSim [10] was used.
Fig. 8 gives a description of IntSim. Fig. 9 reveals that for a
22nm 600MHz logic core, doubling the number of
monolithically stacked device layers can provide similar benefits
to a generation of scaling (2x lower power, 2x lower die size).
These advantages are due to the shorter wires provided by
monolithic 3D, which allow reduced gate (driver) sizes. Note
that Rent’s Rule based stochastic wire length distributions were
used for this analysis, as CAD tools for monolithic 3D design
are still immature today. Conclusions
Several techniques to obtain monolithic 3D chips have been
proposed in this paper. Benefits equivalent to several generations
of scaling can be obtained with the technology without incurring
the cost and risk of next-generation lithography. Wiring and
litho limitations of traditional scaling could make monolithic 3D
increasingly attractive moving forward.
References
[1] W. Dally, Throughput Computing, Intl. Conf. on
Supercomputing, 2010 [2] S. Naffziger, S. on VLSI Technology,
2011 [3] M. Sadaka, et al., Building Blocks for Wafer-Level 3D
Integration, Solid State Technology [4] J. Choi, et al., S. on
VLSI Technology, 2011 [5] R. Liu, Short Course at S. on VLSI
Technology, 2010 [6] H. T. Lue, S. on VLSI Technology, 2010
[7] US Patent # 8,026,521, MonolithIC 3D Inc. [8] Y. Saito, et
al., S. on VLSI Technology, 2000 [9] J. Y. Kim et al., S. on
VLSI Technology, 2003. [10] D. C. Sekar, et al., ICCAD, 2007.
Fig. 1: Exponential increase of litho tool cost.
An immersion litho tool costs $40M today while
other tools in the fab cost $1M-$5M.
Fig. 2: Wire RC delay and energy trends
(Left) nVIDIA 28nm GPUs [1] and (Right) AMD chips [2].
Fig. 5: Comparison with conventional NAND
flash and poly 3D NAND. Fig. 6: Monolithic 3D with conventional logic transistors
would need temperatures higher than 400oC.
Fig. 8: IntSim, an open-source 2D/3D chip simulator. Fig. 9: IntSim’s results for a 22nm 600MHz logic core.
Fig. 3: Ion-Cut process for stacking single
crystal silicon (c-Si) layers at less than 400oC. Fig. 4: Process Flow for monolithic 3D NAND flash with junction-
free charge-trap flash cells made of single crystal silicon.
n Si n Si n Si
H
H
Implant H
Top layer
Bottom layer
Flip top layer
and bond to
bottom layer
400oC anneal or
mechanical force
to cleave. CMP.
Reuse original
substrate.
Use ion-cut to form multilayer
c-Si/SiO2stack above periphery Use shared litho step to define
Form NAND string with
shared litho step
Form contacts and cell source
with shared litho steps
n+ Si Oxide Gate electrode ONO
Cu/Al Cell Source Periphery with W wiring
140 sq.
mm die
2D NAND
@22nm
Poly Vertical 3D
NAND (BiCS)
32 layers @45nm
3D Single Crystal
Silicon NAND
8 layers @22nm
Density
64Gbit
(3 bits/cell)
128Gbit
(1 bit/cell)
256Gbit
(2 bits/cell)
Aspect
ratio
60:1
16:1
Sub-400oC
possible?
Method
Single Crystal Silicon
Yes
Ion-Cut
Shallow Trench Isolation
Yes
Radical Oxidation [8], HDP
High k/Metal Gate
Yes
ALD/CVD
S-D Dopant activation
No
>750oC anneal
Contacts
Yes
Nickel Silicide
Fig. 7: (a) A recessed channel transistor (b) Process flow for monolithic 3D logic. Bottom device layer with Cu/low k does not see more
than 400oC. Through-silicon connections can be close to minimum feature size due to the thin-film process.
Operation
pJ
Integer Add
1
Fetching operands from
A register file
1mm away
26
L1/L2/L3 caches
50/256/
1000
Gate
n+
n+
p
Channel
length tunable
by depth of
recess etch
(a) (b)
Oxide
Activated n+ Si
Activated p Si
Gate Electrode
Gate Dielectric
Cu
Activate dopants
at high temperature
before layer
transfer
Implant
hydrogen
for ion-cut
H
Ion-cut for transfer
of un-patterned c-Si.
No alignment issues.
Finish well-aligned thin-film c-Si recessed
channel transistors @ less than 400oC
Bottom
device
layer
IntSim v2.0
Contains models for:
Stochastic wire length
distributions of 2D/3D-ICs,
logic gates, repeaters, chip
power as well as power,
clock and thermal
interconnect networks
Inputs:
•Gate count
•Die area
•Frequency
•Rent’s
parameters
•Number of
device layers
Outputs:
•Chip
power
•Metal
level
count
•Wire
pitches
10 metal layers per
device layer
2D
@ 22nm
2 layer 3D
@ 22nm
Comment
Average wire length
6um
3.1um
Since 3D and optimal
die size less
Average gate size
6W/L
3W/L
Since less wire cap.
Optimal total silicon
area (Footprint)
50mm2
(50mm2)
24mm2
(12mm2)
Since smaller gates,
shorter wires
Power
1.6W
0.8W
Since gate, wire,
repeater area less