Transcript PJB_UMBC_presentation

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] Tel:301-935-6547
April 4, 2005 UMBC Presentation
What does the future portend for high performance
microelectronic systems for thermal & power distribution?
93W/cm2
40W/cm2
68W/cm2
Think the shuttle has a thermal management problem?
Consider the next generation of high performance chips:
>160 Watts/cm2
State-of-the-Art Thermosyphon in
Commercial Work Stations
• 100 Watt - 6 pipe finned
tower
• 30 cubic inches
• 30 fins per stack
• 100-150 ft3/min air flow
required
• Cools only 1 chip (1cm2 )
Modern Impediments to Chip
Design
• Thermal Limitations < 90 oC Qjc
• Current Density: JAl < 4 x 105Amps/cm2
• Material Breakdown Potential, Emax= Vmax/dmin
Emax (typical) = 1 Volt/10 nanometers = 100 x 106 V/m
• Power Density < 250 Watts/cm2
• Random Soft Errors in Logic & Analog Circuits
– Signals are a countable number of electrons ~ < 105 electrons
– Thermal noise fluctuations
– Quantum noise: particle decay, cosmic rays, radiation induced SEU
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 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.
Moore’s Law
Shannon-von NeumannLandauer (SNL) Switching
Energy Per Bit*
Ebit>ESNL= kBT ln2
= 0.017eV @300K
Using the Uncertainty Relations for x &
t, one calculates the max integration
density of binary switches
~ 5 x 1013 devices/cm2,
Power Dissipation per unit Area
~ 4 x 106 W/cm2 !
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!
* ”Limits to Binary Logic Switch Scaling – A Gedanken Model”, Zhirnov, et al, Proc. of IEEE, Vol. 91,
No. 11, Nov 2003, pp1934-1938.
Spray Cooling:
One Possible Answer
A Phase Change Methodology to Heat
Removal that is Isothermal While Also
Severing the Thermomechanical Heat Path
Phase Change: Spray Cooling System
Can Meet These Requirements
But It Requires An Active Cooling System
Vapor Return
Heat
Rejection
Fan
Condenser/
Heat Exchanger
Low power
density, low DT
heat rejection
Pump Filter
Delivery and Return
Tubes (Can be Part of
Chassis Frame with
Quick-Disconnects)
Heat
Acquisition
Spray Cooling Cap
Cooled MCM
or IC Package
Electronic Circuit Board
Liquid Out
Compact Pump-Condenser Unit
Typically Mounted at Rear of Enclosure or
Externally
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.
Isothermal Spray Cooling
Directly Onto the Chip Surface
Pressurized Liquid
from Pump Enters
Spray Manifold
Liquid
Inlet
Spray Nozzles Create
Fine Cooling Mist
Impinging on IC Chips
20-30 psi
Typical
Vapor and Excess
Spray Drawn Out
Suction Tube for
Return to Condenser
Suction
Outlet
1 atm
(Not Pressurized)
Typical
IC Die
(Shown Flip-Chip
Mounted with Underfill)
MCM Substrate
(Illustrated with Integral
BGA Package)
Printed Circuit
Board
400 Watt Ceramic Package/
MCM with Spray Cooled Lid
400W MCM
Spray Nozzle
Array
Vapor Out
Fluid In
Top View w/Underside of Lid
Side View w/Lid Attached
Phase Change: Spray Cooling
Attributes
•
•
•
•
•
•
•
•
•
•
•
•
Used for point-of-source cooling or system wide cooling
Demonstrated > 150 W/cm2 heat extraction
Forces Isothermal surface conditions wherever phase change occurs
Severs thermomechanical heat path enabling vibration and shock isolation
Direct heat removal from the IC chip surface for minimum Qj
Conformable to surfaces, small liquid volumes required
Phase change fluids enable the choice of operating conditions over a large
temperature range (77K-380K)
Active cooling requires 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 reliability can be 99.999999% with redundancy
Phase change fluids require chemical and mechanical filters for long term
reliability; fluids must be chemically stable, environmentally friendly
Will not meet $0.10/Watt commercial costs in present form
Consider the Materials Used for the
Thermal Conductivity Part of the Problem
Assumptions:
• 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: KDiamond= 20 W/cm C @ 25C, and also,
KDiamond= 100W/cm C @ 100 Kelvin
Polycrystalline Synthetic Diamond
Courtesy of Norton Diamond Film
Electro-Magnetic Arc Deposition of Synthetic Diamond
Visible &Thermal Images of
Ceramic and Diamond MCMs
Diamond Substrate
Ceramic Substrate
False Color IR
Thermal Image @
10 Watts per Chip
Visible Light
Image
Still Air
&
Room
Temperature
Cray Research J90 Supercomputer
MCM Version of a Cray J90
SGI/Cray Research J90 Supercomputer
Rack Mounted Flight Test Version
Monitor
Cray Hard Drive
J90 Cabinet
Module
Power
Supply
Aircraft Interface
All Computers Are Architecturally
Two Dimensional
The first two architectural dimensions are
used for the circuit layout.
The third architectural dimension is used
for heat transfer.
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.
Schematic of
3D Machine
Photo of Sub-nanosecond Cycle Time Machine
Consider High Performance CMOS
Operating at 1 Volt Bias
• A 100 Watt chip requires 100 Amperes
• Electromigration concerns limit current
densities to < 105 A/cm2
• Power distribution with distributed power
converters is only 40 to 60% efficient,
generating even more waste heat
How to Handle the Power
Distribution?
• Why are power converters so inefficient?
– High values of RDS-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
New Silicon NMOS [email protected]
Switch Assembly Top view
• Si NMOS via 2m MOSIS
• 0.4 X 0.4 cm2 chip
• Copper BCB Fine line Package
Interconnect
• 10 micron thick Cu
• C4 Flip Chip Bonded
Side
• 20 mil Cu-Invar-Cu Source
view
Contact Plane
• 17 meter Gate length
Chip
Packaged
Chip
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
Conclusions & Observations
• Spray cooling has demonstrated the capability to handle power
density > 150 W/cm2
• 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
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] Tel:301-935-6547
Laser Embedded Graphite
Resistors in Diamond
1K Ohm
@ 500
Watts
“As Grown” Diamond cut with
a laser showing dark “residue”
around circumference
Laser induced graphite resistors in a diamond
substrate with metal interconnect*
* Boudreaux, et al, “Laser Induced Graphite Resistors in Synthetic Diamond”, International Journal of Microelectronics &
Electronic Packaging, 1996, Volume 19 , pp 169-177.
Moving & Expelling Heat
• Low system overhead
• Isothermal environment
• Qj < 90 oC
• 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
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
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/cm2
Solution: Change Technology to CMOS
• CMOS – ULSI > 100 W/cm2
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.