Paul J. Boudreaux Consider a Mixed Analog/Digital/MEMs

Heat spreaders with high K values attached to the chip can help alleviate lateral DT problems. Placing the system or components into a forced isothermal environment also reduces DT, dT/dt and CTE related problems. Severing the thermomechanical heat path reduces or eliminates shock and vibration from entering the system while reducing weight.

1. Thermal Limitations and
Solutions for Microelectronic
Systems Driven by the
Economics of Moore’s Law
Paul J. Boudreaux
Consultant & Senior Scientist
University Research Foundation
Laboratory for Physical Sciences
College Park, MD 20740
[email protected]:301-935-6547
April 4, 2005 UMBC Presentation

2. Think the shuttle has a thermal management problem?
Consider the next generation of high performance chips:
>160 Watts/cm
2
40W/cm
2
68W/cm
2
93W/cm
2
What does the future portend for high performance
microelectronic systems for thermal & power distribution?

3. State-of-the-Art Thermosyphon in
Commercial Work Stations
•100 Watt -6 pipe finned
tower
•30 cubic inches
•30 fins per stack
•100-150 ft
3
/min air flow
required
•Cools only 1 chip (1cm
2
)

4. Modern Impediments to Chip
Design
•Thermal Limitations < 90
o
C Qjc
•Current Density: J
Al< 4 x 10
5
Amps/cm
2
•Material Breakdown Potential, E
max= V
max/d
min
E
max(typical)= 1 Volt/10 nanometers = 100 x 10
6
V/m
•Power Density < 250 Watts/cm
2
•Random Soft Errors in Logic & Analog Circuits
–Signals are a countable number of electrons ~ < 10
5
electrons
–Thermal noise fluctuations
–Quantum noise: particle decay, cosmic rays, radiation induced SEU

5. Consider a Mixed Analog/Digital/MEMs
System’s In-Plane Thermal Gradients
•dT/dt , DT and CTE are major problems
•Analog voltages are temperature sensitive
•In-situ stress/strain causes failure thru DT
Potential Solutions:
Heat spreaders with high K values attached to the chip can help
alleviate lateral DT problems. Placing the system or components
into a forced isothermal environmentalso reduces DT, dT/dt and
CTE related problems. Severing the thermomechanical heat path
reduces or eliminates shock and vibration from entering the
system while reducing weight.

6. Shannon-von Neumann-
Landauer (SNL) Switching
Energy Per Bit
*
E
bit>E
SNL= k
BT ln2
= 0.017eV@300K
Using the Uncertainty Relations for x &
t, one calculates the max integration
density of binary switches
~ 5 x 10
13
devices/cm
2
,
Power Dissipation per unit Area
~ 4 x 10
6
W/cm
2
!
* ”Limits to Binary Logic Switch Scaling –A Gedanken Model”, Zhirnov, et al, Proc. of IEEE, Vol. 91,
No. 11, Nov 2003, pp1934-1938.
We are at a Thermal Management Barrier(TMB)
long before we get to the limits from Moore’s law
or the semiconductor laws of physics!
Moore’s Law

7. Spray Cooling:
One Possible Answer
A Phase Change Methodology to Heat
Removal that is Isothermal While Also
Severing the Thermomechanical Heat Path

11. Phase Change: Spray Cooling System
Can Meet These Requirements
But It Requires An Active Cooling System
Heat can be acquired at high power density
levels, enabling circuitry to be compacted to
very small board areas, while heat is rejected
remotely, where providing larger condenser
(radiator) areas and airflow is not a problem.
Electronic Circuit Board
Spray Cooling Cap
Heat
Acquisition
Delivery and Return
Tubes (Can be Part of
Chassis Frame with
Quick-Disconnects)
Cooled MCM
or IC Package
Liquid Out
Vapor Return
Fan
PumpFilter
Compact Pump-Condenser Unit
Typically Mounted at Rear of Enclosure or
Externally
Condenser/
Heat Exchanger
Heat
Rejection
Low power
density, low DT
heat rejection

12. Isothermal Spray Cooling
Directly Onto the Chip Surface
Pressurized Liquid
from Pump Enters
Spray Manifold
Spray Nozzles Create
Fine Cooling Mist
Impinging on IC Chips
Vapor and Excess
Spray Drawn Out
Suction Tube for
Return to Condenser
Suction
Outlet
Liquid
Inlet
Printed Circuit
Board
MCM Substrate
(Illustrated with Integral
BGA Package)
IC Die
(Shown Flip-Chip
Mounted with Underfill)
1 atm
(Not Pressurized)
Typical
20-30 psi
Typical

13. 400 Watt Ceramic Package/
MCM with Spray Cooled Lid
Top View w/Underside of Lid Side View w/Lid Attached
Fluid InVapor Out
400W MCM Spray Nozzle
Array

14. Phase Change: Spray Cooling
Attributes
•Used for point-of-source coolingor system wide cooling
•Demonstrated > 150 W/cm
2
heat extraction
•Forces Isothermal surface conditionswherever phase change occurs
•Severs thermomechanical heat pathenabling vibration and shock isolation
•Direct heat removalfrom the IC chip surface for minimum Q
j
•Conformableto surfaces, small liquid volumes required
•Phase change fluids enable the choice of operating conditions over a large
temperature range (77K-380K)
•Active coolingrequires a “dead man’s hand” system to remove trapped heat
within the system when power fails
•Capable of reliable zero-g operation, gives additional cosmic ray protection
from SEU due to electrostatic charge buildup in spacecraft
•Mechanical pump reliabilitycan be 99.999999% with redundancy
•Phase change fluids require chemical and mechanical filtersfor long term
reliability; fluids must be chemically stable, environmentally friendly
•Will not meet $0.10/Wattcommercial costs in present form

15. Consider the Materials Used for the
Thermal Conductivity Part of the Problem
•Simple thermal
conductivity model -
P=dq/dt = K A DT/d
•State-of-the-Art CMOS
•Random logic
•Max clock rate
•3D Interconnected layers
Man-made polycrystalline diamond could passively meet this
requirement for thermal conductivity, K > 13W/cm C.
Note:K
Diamond= 20 W/cm C @ 25C, and also,
K
Diamond= 100W/cm C @ 100 Kelvin
Assumptions:

16. Electro-Magnetic Arc Deposition of Synthetic Diamond
Courtesy of Norton Diamond Film
Polycrystalline Synthetic Diamond

17. Visible &Thermal Images of
Ceramic and Diamond MCMs
Visible Light
Image
False Color IR
Thermal Image @
10 Watts per Chip
Ceramic Substrate
Diamond Substrate
Still Air
&
Room
Temperature

18. Cray Research J90 Supercomputer

20. MCM Version of a Cray J90

21. SGI/Cray Research J90 Supercomputer
Rack Mounted Flight Test Version
Cray Hard Drive
Aircraft Interface
J90 Cabinet
Module
Monitor
Power
Supply

22. All Computers Are Architecturally
Two Dimensional
The first two architecturaldimensions are
used for the circuit layout.
The third architectural dimension is used
for heat transfer.

25. 25KW 3D Diamond Cube Design
Stacked Diamond MCMs Form Cube
Design
Diamond MCM with Cu-Polyimide
Interconnects Layers @ 625 Watts/MCM
Moving the heat laterally through the diamond substrate to the edge fins can produce
reliable operational conditions in this 3D prototype demonstration system
*
at 2.5KW.
Four 3D MCM interconnected substrates make up this prototype system. Spray
cooling (phase change) extracts the heat from the two edge fin areas.
*
“Thermal Analysis of Spray Cooled 3D Interconnected Diamond Substrate MCMs: Comparison With
Experimental Measurements”, Boudreaux, et al, IEEE on-line Journal Transactactions on Device and
Materials Reliability, December 2004, pp594-604.

26. Schematic of
3D Machine
Photo of Sub-nanosecond Cycle Time Machine

27. Consider High Performance CMOS
Operating at 1 Volt Bias
•A 100 Watt chip requires 100 Amperes
•Electromigration concerns limit current
densities to <10
5
A/cm
2
•Power distribution with distributed power
converters is only 40 to 60% efficient,
generating even more waste heat

28. How to Handle the Power
Distribution?
•Why are power converters so inefficient?
–High values of R
DS-on(>20 miliOhms) for semiconductor
switches
–Transformer inefficiency –poor ferrite performance
–Switching power converters use low frequency <150 KHz
•Is there a better way?
–Integrated converters on chip at 100-200MHz
–New semiconductor switch design
–Nanoparticle ferrites for laminar transformers on chip

29. New Silicon NMOS [email protected]
Switch Assembly
•Si NMOS via 2m MOSIS
•0.4 X 0.4 cm
2
chip
•Copper BCB Fine line Package
Interconnect
•10 micron thick Cu
•C4 Flip Chip Bonded
•20 mil Cu-Invar-Cu Source
Contact Plane
•17 meter Gate length
Chip
Packaged
Chip
Top view
Side
view

30. 400 Amperes @ 1.2 Volts CMOS
R
DS(on)=179 microOhms
-2 picosecond “Turn on” time
-CMOS Gate length 17 Meters
-Made with MOSIS technology
-93% efficient power converter
Working prototype of
switching power converter

31. Conclusions & Observations
•Spray cooling has demonstrated the capability to handle power
density > 150 W/cm
2
•Polycrystalline diamond is an excellent CTE matched heat
spreader for Silicon ICs
•Power converters “on-chip” are possible with efficiencies >
90%
•On-chip currents of 200 to 400 Amperes are reliable at 1 volt
•Random errors are a significant problem:
–Redundancy (voting) is often not a viable solution with limited power
–New “random error hardened” logic designs are required
–Analog designs are now limited by these random level signals

33. A Case Study in Thermal Management
and Power Distribution For a 3D
Interconnected High Performance
Microelectronic System
Paul J. Boudreaux
Consultant & Senior Scientist
University Research Foundation
Laboratory for Physical Sciences
College Park, MD 20740
[email protected]:301-935-6547

34. Laser Embedded Graphite
Resistors in Diamond
“As Grown” Diamond cut with
a laser showing dark “residue”
around circumference
Laser induced graphite resistors in a diamond
substrate with metal interconnect
*
1K Ohm
@ 500
Watts
* Boudreaux, et al, “Laser Induced Graphite Resistors in Synthetic Diamond”, International Journal of Microelectronics &
Electronic Packaging, 1996, Volume 19 , pp 169-177.

35. Moving & Expelling Heat
•Low system overhead
•Isothermal environment
•Q
j< 90
o
C
•Die attach thermal impedance
minimized
•3D interconnections allowed
•Clock speeds > 4GHz
•Bias voltage ~ 1 Volt
•High thermal conductivity
materials
•CTE matching materials
•High (>90%) efficiency power
supplies
•Phase change heat removal

36. WHY?
The “best” switching supplies are limited
by the switching device’s RDS-on
impedance, typically 10 –20 miliOhms.
What if a switching transistor were made
with a gate length of 50 meters instead of 2
–3 microns? Then RDS-on would be < 20
microOhms!At a clock speed of 200 MHz
instead of 100 kHz, the L-C-R components
are fully embedded and Integrated into die

37. déjà vu
This is not the first time the industry has
run into TMB
•Vacuum tube –ENIAC’s MTBF =15min Solution:
Change Technology to transistors
•Bipolar semiconductors –TTL, ECL, LSI~100W/cm
2
Solution: Change Technology to CMOS
•CMOS –ULSI > 100 W/cm
2
Solution: ? Today we are better situated with new
materials and cooling technology to deal with the problem.
NOTE: Even if CMOS is not used, the thermal problem remains, i.e. SNL data.
3D structures introduce a paradigm shift in thermal design because prior
systems were architecturally two dimensional; the third dimension was the
surface where the heat was extracted. In a true 3D environment, this third
dimension is used by interconnects, power distribution and bonding. New
thermal concepts will be required to extract heat from such 3D systems.