Transcript Slide 1
3D FPGA
and Trends of Future Logic Fabric
Zvi Or-Bach
MonolithIC 3D
President & CEO
MonolithIC 3D Inc. Patents Pending
Agenda
Scaling and the Logic Challenge
Current FPGAs
The Third Dimension
The Architecture Opportunity
The Technology Opportunity
The Opportunity for Structured ASIC
Summary
MonolithIC 3D Inc. Patents Pending
Transistors no Longer Dominate – Metal Interconnections Took Over
MonolithIC 3D Inc. Patents Pending
Interconnects Dominate with Scaling [Source: ITRS]
90nm (2005)
45nm (2010)
22nm (2015)
12nm (2020)
1.6ps
0.8ps
0.4ps
0.2ps
Delay of 1mm long
Interconnect
5x102ps
2x103 ps
1x104 ps
6x104 ps
Ratio
3x102
3x103
4x104
3x105
Transistor Delay
Transistors keep improving
Surface scattering, grain boundary scattering and diffusion
barrier degrade RC delay
Low k helps, but not enough to change trend
MonolithIC 3D Inc. Patents Pending
4
MonolithIC 3D Inc. Patents Pending
Accelerating Mask-Set Cost
Increase of interconnection complexity
Increase in mask complexity
Process (m)
=> More Metal Layers
=> Mask cost increases
2.0 … 0.8 0.6 0.35 0.25 0.18 0.13 0.09
Single Mask
1.5
cost ($K)
1.5 2.5 4.5 7.5 12
40
60
# of Masks
12
12
30
34
Mask Set
cost ($K)
18
18 30 72 150 312 1,000 2,000
12
16
20
MonolithIC 3D Inc. Patents Pending
26
IC Design Costs
26
24
22
20
18
16
14
12
10
8
6
4
2
0
0.35 micron
(feature dimensions vs. cost in $ millions)
Prototype
Validation
Physical
Verification
Architecture
0.25 micron
0.18 micron
0.13 micron
90 nm
Source: International Business Strategies
MonolithIC 3D Inc. Patents Pending
Number of Design Starts is Collapsing
Design Starts
Source: VLSI Research, Inc.
MonolithIC 3D Inc. Patents Pending
20
14
20
12
20
10
20
08
20
06
20
04
20
02
20
00
19
98
19
96
19
94
19
92
45,000
40,000
35,000
30,000
25,000
20,000
15,000
10,000
5,000
0
Semiconductors Market Growth
MonolithIC 3D Inc. Patents Pending
(per Quarter)
MonolithIC 3D Inc. Patents Pending
FPGAs Aren't Substituting ASICs
Total FPGA Market <$5B, Slow Growth Since 2000
FPGA Market
5000
4500
4000
$, Millions
3500
3000
Actel
2500
Altera
2000
Xilinx
1500
1000
500
0
1999
2000
2001
2002
2003
2004
2005
2006
Calendar year
MonolithIC 3D Inc. Patents Pending
2007
2008
2009
2010
The Logic Market Growth
MonolithIC 3D Inc. Patents Pending
FPGA at ~$4B Clearly didn’t Fill the ASIC Gap
FPGA market has barely grown in the last 5 years
FPGA cost-speed-power make it unattractive for many volume
applications
FPGA penalty is extremely high *
Gate Density: 1: 20-40
Power:
1: 9-12
Speed:
1: 2-4
Old Process ASIC is preferable in many applications
Density ~1:32=> 5 process generations (45nm FPGA ~ 0.25m ASIC)
=> 3-4 gen. older ASIC process (0.13m-0.18m) is more competitive
=> 3-4 gen. older ASIC process fab is fully deprecated and hence cheaper
ASSP + Software in many cases provide a better alternative
*Ian Kuon and Jonathan Rose, "Measuring the Gap Between FPGAs and ASICs", IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems (TCAD) , Vol. 26, No. 2, pp 203-215, Feb. 2007
MonolithIC 3D Inc. Patents Pending
FPGA Achilles’ Heel – PIC (Programmable
Interconnect) >30x Area vs. Masked Via
SRAM
bit
SRAM FPGA connectivity
elements @ 45 nm
Via connectivity element
@ 45 nm
SRAM
bit
4X
4X
Bidi buffer
Via/AF
Via
Area 4 m2
Ratio to AF 100
.2m
Area: 0.04 m2
SRAM
bit
.2m
TS buffer
4X
Area 2 m2
Ratio to AF 50
SRAM
bit
Pass gate
10X
Pitch - 0.2 m
.2 x .2= 0.04 m2
Area .5 m2
Ratio to AF 12
Current FPGAs use
primarily pass
transistors with a driver.
Average area ratio of connectivity element >30
MonolithIC 3D Inc. Patents Pending
FPGAs see Diminishing Benefits with Scaling
Over 90% of FPGA programmable logic area penalty is due to
programmable interconnect (‘PIC’)
Performance and power penalty are direct result of the area
Transistor as Programmable Interconnect doesn't scale well
Scaling bring on “The Tyranny of Interconnects”
Interconnect needs to increase faster than gate count to keep up (Rent’s
rule)
ASIC adds metal layers with scaling
PIC limiting factor is the single layer of diffusion shared with the logic.
=> FPGA moved from LUT4 to LUT6-7 as higher
granularity logic cells reduce interconnections
MonolithIC 3D Inc. Patents Pending
Current 3D FPGA Status
Tier Logic
Closed Doors
Use TFT to build the Programming RAM on top – 2nd tier
Suggested cost reduction for high volume by removing the TFT
layers and replacing the PIC with ‘Masked Link’
Tabula
3D of a different kind – Time
Just closed additional round of $100M – we wish them luck
NuPGA
Just changed its name to MonolithIC 3D
Antifuse 3D FPGA Concept.
On the way to a better FPGA discovered a path for practical
monolithic 3D and shifted its strategy to focus on it.
MonolithIC 3D Inc. Patents Pending
Other non-commercial ideas
Placing configuration
SRAM at different
strata (Lin, El Gamal,
et al., Stanford 20062007)
Placing hard blocks
like DSP & user
memory at different
strata (Le, Reda &
Bahar, Brown U.
2009)
RR – Routing Resource
LB – Logic Block
ST – Switch Transistor
(pass gate)
LB-SRAM – LB configuration
memory
RR-SRAM – RR configuration
memory
MonolithIC 3D Inc. Patents Pending
Defect Density
Xilinx 2.5D
MonolithIC 3D Inc. Patents Pending
Altera 3D
MonolithIC 3D Inc. Patents Pending
The Future - 3D IC
1950s
Today
Too many interconnects to manually solder
interconnect problem
Interconnects dominate performance and
power and diminish scaling advantages
interconnect problem
Solution: The (2D) integrated circuit
Solution: The 3D integrated circuit
Kilby version:
Connections not integrated
3D with TSV: TSV-3D IC
Connections not integrated
Noyce version
(the monolithic idea):
Connections integrated
Monolithic 3D: Nu-3D IC
Connections integrated
MonolithIC 3D Inc. Patents Pending
20
FPGAs:
The 3D-TSV Solution
MonolithIC 3D Inc. Patents Pending
Current 3D-IC : TSV
MonolithIC 3D Inc. Patents Pending
Reinventing FPGA using 3D-TSV
Use different dies for:
Programmable Logic
Programmable I/O
Programmable Memory
Build Reticle Size with Multi-Dice Lines for each
~One Mask set with many die/functions size options
Choice of process node and fabrication processes
MonolithIC 3D Inc. Patents Pending
Traditional FPGA Wafers vs. Continuous Array Wafers
FPGA
chip with
IO
Scribe
lanes
Scribe
lanes
IO “chiclet”
with TSV
prep
Traditional wafer of chips
TSVs
FPGA
Logic-only
chunk
TSV
Continuous Array of Logic at 22nm
Wafer of IO Chiclets at 0.15 m
MonolithIC 3D Inc. Patents Pending
Continuous Array Terrain Allows Defining Custom
Logic Sizes from Same Wafers
Chip
size 9
Chip
size 4
Scribe
lane
Chip
size 20
FPGA
Logic-only
chunk
Long metal tracks cross scribe lines
Die edges need to be sealed after cut
Top wafer prepped for TSVs or micorbumps in standard pattern (“socket”) over logic
chunks
MonolithIC 3D Inc. Patents Pending
Assembling Continuous Array Terrain into
Customized Hybrid Stacks
Standard TSV
pattern
Chip
with
logic
size 9
Chip with TSV
visible
3D Hybrid
stack with
TSVs
TSV
prep
IO
0.15 m
memory
22 nm
SerDes
90 nm
• Chiclets
assembly actually happens prior to wafer
dicing
Chiclets
MonolithIC 3D Inc. Patents Pending
Advantages of 3D-TSV FPGA
Good Fit for the End User Application with optimal size of
silicon - ~2-5 x cost reduction
Wide range of technologies and function
Reduce cost by using low cost old process for I/O
Increase functionality to match user alternative SC
Allow integration of additional vendors with their own dies
Huge reduction of $$$ for masks
MonolithIC 3D Inc. Patents Pending
Future SoC New Logic 15%
MonolithIC 3D Inc. Patents Pending
FPGAs:
The Monolithic 3D Solution
MonolithIC 3D Inc. Patents Pending
3D IC Next Generation – Monolithic 3D
Monolithic 3D vs. TSV
(1:10,000 vertical connectivity ratio)
TSV: TSV-3D IC
Monolithic: Nu-3D IC
Layer Thickness
~50m
50-100nm
Via Diameter
~5m
~50nm
Via Pitch
~10m
~100nm
Limiting Factors
Wafer handling (~5m)
Aspect ratio (<10:1)
Lithography=>
Will keep scaling
Wafer (Die) to Wafer
Alignment
~1m
Layer to Layer Alignment
MonolithIC 3D Inc. Patents Pending
=> Will
keep scaling
Layer Transfer Technology (“Ion-Cut”)
Defect-free single crystal formed @ <400oC
Oxide
Hydrogen implant
Flip top layer and
Cleave using 400oC
of top layer
bond to bottom layer
anneal or sideways
mechanical force. CMP.
p- Si
Top layer
Oxide
p- Si
Oxide
Bottom layer
p- Si
p- Si
H
Oxide
Oxide
Similar process (bulk-to-bulk) used for manufacturing all
SOI wafers today
MonolithIC 3D Inc. Patents Pending
31
Monolithic 3D with State of the Art Transistors
Uses a novel combination of four ideas
Gate-Last Process and proper sequence of “Ion-Cut”
Low Temperature Face-up Layer Transfer
Repeating Layouts
Innovative Alignment
MonolithIC 3D Inc. Patents Pending
32
A Gate-Last Process for Cleave and Layer Transfer
NMOS
PMOS
Poly
Oxide
Donor wafer
Fully constructed transistors attached to
each other; no blanket films.
proprietary methods align top layer atop
bottom layer
Device wafer
MonolithIC 3D Inc. Patents Pending
33
A Gate-Last Process for Cleave and Layer Transfer
Poly
Oxide
Step 1. On donor wafer,
fabricate standard
dummy gates with oxide,
poly-Si
S/D Implant
Step 2: Std Gate-Last
Self-aligned S/D implants
Self-aligned SiGe S/D
High-temp anneal
Salicide/contact etch stop
or faceted S/D
Deposit and polish ILD
ILD
MonolithIC 3D Inc. Patents Pending
CMP to top of
dummy gates
34
A Gate-Last Process for Cleave and Layer Transfer
Step 3.
Implant H for cleaving
H+ Implant Cleave Line
Step 4.
Bond to temporary carrier wafer
(adhesive or oxide-to-oxide)
Cleave along cut line
CMP to STI
Carrier
STI
CMP to STI
MonolithIC 3D Inc. Patents Pending
35
A Gate-Last Process for Cleave and Layer Transfer
Step 5.
Low-temp oxide deposition
Bond to bottom layer
Remove carrier
Oxide-oxide bond
Remove (etch) dummy
gates, replace with HKMG
Step 6. On transferred layer:
Etch dummy gates
Deposit gate dielectric and electrode
CMP
Etch tier-to-tier vias thru STI
Fabricate BEOL interconnect
MonolithIC 3D Inc. Patents Pending
36
Future Monolithic 3D FPGA
Multi-Tier Programmable Logic
Tier 0 – The LUT Array + Local programmable interconnect
Tier 1 – The Clock distribution Network + programmable power
distribution
Tier 2 – Short Programmable Interconnect
Tier 3 – Long Programmable Interconnect
Reinvented PIC
Antifuse
1T Memory cell instead of the 6T
Flash
DRAM
…
MonolithIC 3D Inc. Patents Pending
Summary
Interconnects are now dominating all logic devices
Early innovations of 3D FPGA stimulate more ideas
FPGA vendors are moving into 3D (so far 2.5D)
Future FPGAs will utilize 3D technology to Reinvent the
FPGA
TSV to re-architect the FPGA system
Monolithic to re-architect the programmable logic fabric
The Future is in the Third Dimension – 3D
MonolithIC 3D Inc. Patents Pending