Challenges and Opportunities in Gen3 Embedded
Cooling with High-Quality Microgap Flow
Avram Bar-Cohen
1
, Franklin L. Robinson
2
, and David C. Deisenroth
3
1
Space and Airborne Systems, Raytheon Company, Arlington, VA 22209, USA
2
Thermal Engineering Branch, NASA Goddard Space Flight Center, Greenbelt, MD 20771, USA
3
Department of Mechanical Engineering, University of Maryland, College Park, MD 20742, USA
Abstract—Gen3, Embedded Cooling, promises to revolutionize
thermal management of advanced microelectronic systems by
eliminating the sequential conductive and interfacial thermal
resistances which dominate the present “remote cooling”
paradigm. Single-phase interchip microfluidic flow with high
thermal conductivity chips and substrates has been used
successfully to cool single transistors dissipating more than
40kW/cm
2
, but efficient heat removal from transistor arrays,
larger chips, and chip stacks operating at these prodigious heat
fluxes would require the use of high vapor fraction (quality), two-
phase cooling in intra- and inter-chip microgap channels. The
motivation, as well as the challenges and opportunities associated
with evaporative embedded cooling in realistic form factors, is the
focus of this paper. The paper will begin with a brief review of the
history of thermal packaging, reflecting the 70-year “inward
migration” of cooling technology from the computer-room, to the
rack, and then to the single chip and multichip module with
“remote” or attached air- and liquid-cooled coldplates. Discussion
of the limitations of this approach and recent results from single-
phase embedded cooling will follow. This will set the stage for
discussion of the development challenges associated with
application of this Gen3 thermal management paradigm to
commercial semiconductor hardware, including dealing with the
effects of channel length, orientation, and manifold-driven
centrifugal acceleration on the governing behavior.
Keywords—embedded cooling; electronics cooling; two-phase
flow; heat transfer; microgap; microchannel
I. STATE-OF-THE-ART IN EMBEDDED COOLING
The increased integration density of electronic components
and subsystems, including the nascent commercialization of 3D
chip stack technology, has exacerbated the thermal management
challenges facing electronic system developers. The confluence
of chip power dissipation above 100 W, localized hot spots with
fluxes above 1 kW/cm
2
, and package-level volumetric heat
generation that can exceed 1 kW/cm
3
has exposed the limitations
of the current “remote cooling” paradigm and its inability to
support continued enhancements in the performance of
advanced silicon and compound semiconductor components.
These thermal limitations have compromised the decades-long
Moore’s law progression in microprocessor performance and
threaten to derail the innovation engine which has been
responsible for much of the microelectronic revolution [1,2].
In conventional cooling architectures for electronics,
reliance on thermal conduction and spreading, in the commonly-
used chips and substrates and across the multiple material
interfaces present in packages and modules, severely constrains
the ability of remotely located heat rejection surfaces to reduce
the temperature rise of critical on-chip hot spots and individual
chips in a module or in a stack. Moreover, continued application
of this “remote cooling” paradigm has resulted in advanced
electronic systems in which the thermal management hardware
accounts for a large fraction of the volume, weight, and cost, and
undermines efforts to transfer emerging electronic components
to portable, as well as other small form-factor, applications.
Consequently, many commercial and military electronic
systems are today “thermally-limited,” i.e., performing well
below the inherent electrical capability of the device technology
they exploit. To overcome these limitations and remove a
significant barrier to continued Moore’s law progression in
electronic components and systems, it is essential to “embed”
aggressive thermal management in the chip, substrate, and/or
package and directly cool the heat generation sites. The
development of such “Gen3” thermal management technology,
following on the Gen1 air-conditioning approaches of the early
years and the decades-long commitment to the Gen2 “remote
cooling” paradigm, has been spearheaded by the Defense
Advanced Research Projects Agency (DARPA) [3-5] and
embraced by many organizations [2,6-8].
The Near Junction Thermal Transport (NJTT) program,
initiated in 2011, was the first program to develop thermal
management for the region within 100 μm of the electrical
junction of a gallium nitride (GaN) transistor and aimed to
enable heat fluxes of greater than 100 W/mm
2
(10 kW/cm
2
)
while maintaining reliable junction temperatures. Through
technology developed in this program, specifically the
placement of GaN epitaxy on—or in close proximity to—high
thermal conductivity diamond, the RF power handling
capability of GaN high-electron-mobility transistor (HEMT)
devices was increased by greater than a factor of 3 [9-14], to in
excess of 500W/mm
2
of RF power and more than 400W/mm
2
of
heat, as shown in Fig. 1 and described in greater detail in [8,15].
While the NJTT program made significant gains using high
thermal conductivity substrates to spread the heat close to the
junction, it did not address the next link in the thermal resistance
chain, i.e. extracting the heat from the diamond and transferring
that dissipated heat to an available coolant. In the Computational
track of the DARPA ICECool Applications program, IBM—
which had earlier demonstrated an ability to remove nearly
4 kW/cm
3
from a chip stack with inter-chip (or interlayer) fluid
flow [16]—used the evaporation of a refrigerant flowing radially
across a single chip to successfully control the temperature of a
2 kW/cm
2
hot spot in cores dissipating greater than 350 W/cm
2
on a Power PC7 chip [17,18]. Georgia Tech, working with Intel-
Altera to apply liquid cooling to a field-programmable gate array
(FPGA) chip, was able to demonstrate a significant increase
(~1.5x) in computational throughput for the Stratix5 by using
forced flow across the back of the chip to reduce leakage
currents and simultaneously cooling all 9 cores on the chip [19].
Fig. 1: Improvement achieved by NJTT performers in HEMT RF areal
power density, the RF output power over the footprint of the HEMT, as a
result of enhanced thermal management enabling operation at higher
transistor RF heat fluxes, the heat dissipated over the footprint of the
HEMT, adapted from [7]
The Power Amplifier track of the DARPA ICECool
Applications program, which began in 2013 [5] and is now
nearing completion, combined embedded single-phase
microfluidics with high thermal conductivity substrates to
reduce the thermal resistances in the entire monolithic
microwave integrated circuit (MMIC) package. The successful
demonstration of embedded microfluidic cooling of the MMIC
substrates with bonded, liquid manifolds, by all participating
research teams, is displayed in Fig. 2. It may be noted that these
results offer 1 kW/cm
2
cooling for the entire chip
(approximately 5 mm by 5 mm) and record-level HEMT
transistor heat removal rates of more than 30 kW/cm
2
or
300 W/mm
2
[15].
Fig. 2: ICECool Applications Phase 1 TDV Results [15]
The dramatic improvements in demonstrated single-
transistor heat removal rates provide a most compelling
validation of Embedded Cooling and set the stage for
commercial and military applications of this Gen3 thermal
management paradigm. However, if such efficient kW-level
heat removal is to be extended from single transistors to
transistor arrays, larger chips, and chip stacks, it will be
necessary to fully exploit the latent heat of evaporation of the
flowing liquid through two-phase local cooling and high exit
vapor fractions. To harness such high-quality, vigorous phase-
change processes in embedded microchannels, cut through the
chips and/or in inter-chip microgap channels, requires a
fundamental understanding of the underlying hydrodynamic
behavior and thermal characteristics of two-phase flow in
manifold-supplied chip-scale coolers operating in a variety of
orientations. Subsequent sections will identify the challenges
implicit in widespread application of embedded two-phase
cooling and provide recent results from the authors’ work in
such configurations [20-24].
II. THERMOFLUID CHARACTERISTICS OF TWO-PHASE
LABORATORY MICROGAP CHANNELS
In two-phase flow, the extent, aggregation, and distribution
of each phase is classified into distinct categories called flow
regimes. The dominant regimes in two-phase microgap flow are
bubbly, intermittent, and annular flow [25]. Bubbly flow is
observed at the low qualities and consists of small, spherical
bubbles, dispersed and transported by the continuous liquid
phase. With increasing flow quality, the bubbles grow and
coalesce until eventually confined by the lateral bounds of the
channel and then continue to elongate axially. This distribution
of large vapor slugs, separated by liquid plugs or bridges, is
known as intermittent flow. As the vapor fraction continues to
increase, the larger vapor bubbles coalesce into a continuous
annular configuration, consisting of a high-velocity vapor core,
and thin, shear-driven, liquid film around the perimeter of the
channel. These flow regimes can be clearly seen in Fig. 3 for the
two-phase flow of R134a in a 0.79 mm diameter channel.
Fig. 3: Prevailing flow regimes in microscale channels for R134a flowing
at 500 kg/m
2
-s in a 0.79 mm diameter tube: (a) bubbly flow at x = 0, (b)
intermittent flow at x = 0.11, and (c) annular flow at x = 0.73 [26]
Recent studies [27-33] of two-phase thermofluid behavior in
relatively long microgap channels have
uncovered a strong dependence of the previously observed M-
shape variation in heat transfer coefficient, as seen in Fig. 4, on
the prevailing two-phase flow regimes. The low-quality peak in
the M-shape heat transfer coefficient profile corresponds to the
incipience of nucleate boiling or bubbly flow. The bubble
nucleation, movement, and acceleration disrupt and thin the
thermal boundary layer, all of which enhance the heat transfer
coefficient. With the additional increase in quality, the flow
quickly progresses from bubbly to intermittent flow. In
intermittent flow, the thermal transport enhancement attributed
to bubbly flow is gradually suppressed by the confinement of
large vapor bubbles and decreasing liquid plug length, resulting
in an overall deterioration in the heat transfer coefficient with
increasing flow quality. However, this deterioration is soon
overcome by the transition from intermittent to annular flow and
driven by the increasing vapor-liquid velocity difference and
thinning of the evaporating liquid film, resulting in an inflection
and monotonic increase in the heat transfer coefficient, until
reaching a second maximum and then followed by a
deterioration that asymptotically approaches the value for
single-phase vapor convection. The specific mechanisms
responsible for this high-quality peak and deterioration in the
heat transfer coefficient are poorly understood and have
constrained the design, optimization, and implementation of
two-phase coolers in the favorable high-quality domain.
Fig. 4: Characteristic M-shape variation in the heat transfer coefficient
for two-phase refrigerant flow in microgap channels. Data from [31-33].
In an investigation of the mechanisms responsible for the
thermal deterioration observed in high-quality annular flow,
Kabov, Zaitsev, Cheverda, and Bar-Cohen [34] empirically
simulated an adiabatic shear-driven liquid film in a 40 mm by
80 mm by 2 mm channel by independently injecting liquid
FC-72 and gaseous nitrogen streams. Deformations and
emerging patterns were observed at the liquid-vapor interface
and, depending on the flow rate of each phase, were classified
into five sub-regimes: (1) cells, (2) structures, (3) 2D waves, (4)
3D waves, and (5) film rupture. The sub-regime map of the
interfacial deformations and a photographic example of the 2D
and 3D wave patterns are shown in Fig. 5. At low liquid and
vapor velocities, the deformations in the liquid film are weak
and classified as cells and structures. As the vapor velocity
increases, periodic 2D waves emerge on the surface of the liquid
film; the frequency of the 2D waves increases with vapor
velocity until eventually evolving into a 3D wave structure. For
low liquid film flow rates and sufficiently large vapor velocities,
the liquid film ruptures and is torn from surface. The observed
deformations and rupture of the shear-driven liquid film may be
responsible for the thermal deterioration at high vapor qualities
in diabatic microgap channels, as observed and reported in
[20,21,35].
Fig. 5: Sub-regime map and visualization of interfacial deformations for
two-phase flow of FC-72 and nitrogen in a mini-channel, where: (1) cells,
(2) structures, (3) 2D waves, (4) 3D waves, and (5) film rupture [34]
III. CHIP-SCALE EFFECTS IN MICROGAP COOLERS
In applying microgap two-phase cooling to operational
electronic equipment, attention needs to be focused on relatively
short, low length-to-diameter ratio () “chip-scale”
configurations, operating at moderate to high heat fluxes and
high qualities; these differ considerably from the conditions
studied in much of the literature. Recently published studies
[26,29,36,37] have shed considerable light on the characteristics
of such chip-scale microgap channels, but these initial results are
insufficient to establish the axial progression in flow regimes
and the resulting variation (M-shaped or otherwise) of the local
heat transfer coefficient. Consequently, Bar-Cohen and
Holloway [21] undertook an effort to provide an in-depth
exploration of the two-phase thermofluid behavior of a chip-
scale microgap channel using both photographic and infrared
temperature visualizations. The photographic visualization was
used to identify prevailing flow regimes, while the infrared
visualization targeted the associated spatial and temporal wall
temperature variations and their associated heat transfer
coefficients.
In the Bar-Cohen and Holloway study [21], FC-72 was
pumped through a 10 mm by 12 mm by 0.2 mm microgap
channel at a constant flow rate of 0.25 ml/s (corresponding to a
mass flux of 210 kg/m
2
-s) with an inlet subcooling of 5 K. Heat
was applied uniformly to the bottom of the microgap channel at
a constant rate of 7.5, 15.0, 22.5, or 30W, corresponding to a
heat flux of 2.8, 7.9, 13.1, or 17.7 W/cm
2
, and exit quality of
0.09, 0.27, 0.45, or 0.61. This set of conditions was chosen in
order to span a wide range of flow qualities and regimes while
remaining below the temperature limit of the resistance heater
used in the microgap test section (100°C). The surface
temperatures measured by the infrared camera were used to
calculate the time-averaged heat flux distribution, which in turn
was used to calculate the heat transfer coefficient distribution at
the wetted microgap surface. Photographic visualization was used
to evaluate the dominant flow regimes for each set of conditions
and compared to the thermal data to determine any
interdependence between the two.
A schematic of the experimental flow loop used in this study
is shown in Fig. 6 and described in detail in [21]. A top-down,
cross-sectional and exploded view of the assembled copper
microgap test section is shown in Fig. 7. There are two small
ridges on each side of the microgap, used to set the channel gap
and support the sapphire window which confines the upper
surface of the microgap channel and is transparent in both the
visual and midwave infrared spectrums. Heat is applied to the
bottom of the microgap pedestal with a 12 mm by 12 mm
ceramic resistive heater. Five vertical holes were drilled into the
underside of the microgap channel, along the centerline, ending
1 mm below the microgap surface, to allow placement of E-type
thermocouples to measure the centerline temperature of the
microgap heated wall. The microgap surface is coated with a
thin layer of Aeroglaze Z307 Polyurethane Coating to provide a
high and uniform emissivity for infrared temperature
measurements. The paint is 38 µm (0.0015”) thick and has a
thermal conductivity of 0.35 W/m-K.
The synchronized photographic and infrared visualization
for a mass flux of 210 kg/m
2
-s is shown in Fig. 8 through Fig.
11 for an exit quality of 9%, 27%, 45%, and 61%, respectively.
For each case, two sets of visualization—captured at different
times, with the first (left) image reflecting “typical” conditions
and the second (right) image reflecting occasional or atypical
behavior—are presented to demonstrate the unsteady nature of
the flow. The flow is from left to right for all images presented.
Fig. 6: Schematic of flow loop apparatus [21]
Fig. 7: Test section assembly for microgap channel, top, cross sectional
view (left) and exploded view (right) [21]
At a heat flux of 2.8 W/cm
2
and an exit quality of 9%, bubbly
flow is the dominant flow regime, as shown in Fig. 8(a).
However, as shown in Fig. 8(b), large, confined vapor bubbles
are occasionally observed to consume most of the channel and,
in doing so, induce a brief 2 to 3 K increase in wall temperature.
(a)
(b)
Fig. 8: Instantaneous photographic and infrared visualization of two-
phase diabatic microgap flow for FC-72, q’’ = 2.8 W/cm
2
, x
exit
= 9%, h
avg
= 4300 W/m
2
-K, G = 210 kg/m
2
-s, ∆T
sub
= 5 K, and a 10 mm by 12 mm by
0.2 mm channel (a) Typical and (b) Occasional. Flow is left to right.
As the heat flux increases to 7.9 W/cm
2
and the exit quality to
27%, bubbly flow persists at the entrance of the channel but the
vapor bubbles grow and coalesce at a considerably faster rate, as
shown in Fig. 9(a). At this point, the flow for the second half of
the channel consists of very large, confined vapor bubbles that
occasionally (Fig. 9(b)) grow to fill much of the channel, and
may be classified as “churning” intermittent flow. This
“churning” flow of confined bubbles appears to initiate periodic
dryout of the liquid film on the heated wall, producing more
significant, 8 to10 K, wall temperature fluctuations.
As the heat flux continues to 15.7 W/cm
2
and the exit quality
to 45%, periodic intermittent flow dominates most of the
microgap channel and bubbly flow is limited, on average, to
only the first few millimeters of the channel length, as shown in
Fig. 10(a). In the lower left-hand corner of the photographic
image shown in Fig. 10(a), local dryout is observed to occur
along the bottom of the vapor bubbles and periodically grows to
consume a large portion of the channel, oscillating vigorously
between the conditions shown in Fig. 10(a) and Fig. 10(b),
respectively. Under the influence of the higher wall heat flux,
the average wall “excess” temperature increases to
approximately 25 K above the inlet liquid temperature of 51°C,
but more vigorous churning flow limits the local wall
temperature fluctuations to 8 to 10 K, the same range as the
previous heat flux condition.
(a)
(b)
Fig. 9: Instantaneous photographic and infrared visualization of two-
phase diabatic microgap flow for FC-72, q’’ = 7.9 W/cm
2
, x
exit
= 27%, h
avg
= 4500 W/m
2
K, G = 210 kg/m
2
-s, ∆T
sub
= 5 K, and a 10 mm by 12 mm by
0.2 mm channel (a) Typical and (b) Occasional. Flow is left to right.
(a)
(b)
Fig. 10: Instantaneous photographic and infrared visualization of two-
phase diabatic microgap flow for FC-72, q’’ = 15.7 W/cm
2
, x
exit
= 45%,
h
avg
= 5000 W/m
2
K, G = 210 kg/m
2
-s, ∆T
sub
= 5 K, and a 10 mm by 12 mm
by 0.2 mm channel (a) Typical and (b) Occasional. Flow is left to right.
For the highest heat flux of 17.7 W/cm
2
and exit quality of
61%, bubbly and intermittent flow is still observed in the first
half of the channel, though at relatively high wall excess
temperatures of 40 K. But—despite the pulsatile flow in the
channel and the relatively high liquid fraction available
(approximately 40%)—the flow appears not to wet the second
half of the channel, leading to a broad area of elevated wall
temperature, reaching a superheat of more than 45 K. While thin
film annular flow could have been expected under these flow
conditions (based on flow regime predictions), the observed
elevated temperatures and the visual imagery in Fig. 11(a) are
indicative of wall dryout in the second half of the channel, which
grows, periodically, to cover the majority of the channel, as
shown in Fig. 11(b). Interestingly, under these conditions, 3D
waves, rivulets, and hole formations are observed in the shear-
driven film on the upper channel surface (the sapphire window),
which is only weakly heated by conduction from the heated
copper test section into the upper sapphire window.
(a)
(b)
Fig. 11: Instantaneous photographic and infrared visualization of two-
phase diabatic microgap flow for FC-72, q’’ = 17.7 W/cm
2
, x
exit
= 61%,
h
avg
= 4800 W/m
2
K, G = 210 kg/m
2
-s, ∆T
sub
= 5 K, and a 10 mm by 12 mm
by 0.2 mm channel (a) Typical and (b) Occasional. Flow is left to right.
The stability and integrity of the liquid-vapor interface in
two-phase flow results from the interplay between stabilizing
(gravity, surface tension, and viscous damping) and
destabilizing (inertia, mass loss, evaporative recoil, and
thermocapillary) forces. From analyzing the visualization
shown in Fig. 8 to Fig. 11, it would appear that the lower liquid
film is periodically rupturing and dewetting the heated surface,
resulting in a subsequent rise in the local wall temperature. It is,
however, unclear which destabilizing mechanisms are primarily
responsible for this periodic occurrence of film rupture and
elevated wall temperatures. Findings from Kabov, Lyulin,
Marchuk, and Zaitsev [38] and Gatapova and Kabov [39] have
shown that nonuniformities in surface temperature can lead to
significant thermocapillary forces that in turn induce spatial
variations in the film’s thickness. Regions of localized thinning
(negative temperature gradients) are thus especially susceptible
to film rupture. For elevated heat fluxes and qualities, it also
appears that, in a manifestation of a Leidenfrost-like behavior,
the available liquid in the flow is unable to migrate against the
flow of generated vapor and rewet the heated surface.
The average heat transfer coefficients for the conditions
examined in this study are plotted in Fig. 12, above the flow
regime maps, as a function of average superficial vapor velocity,
along with the Chen [40] and Shah [41,42] correlations. The
spatially and temporally averaged microgap heat transfer
coefficient varied only slightly from 4300 W/m
2
-K at an exit
quality of 9%, to a peak of 5000 W/m
2
-K at an exit quality of
45%, and to a slightly lower value of 4800 W/m
2
-K at the
highest exit quality of 61%. While long microgap channels
display an M-shaped variation in the heat transfer coefficient
and both the Chen and Shah correlations predict that the heat
transfer coefficient will rise with increasing quality over the
![Fig. 4: Characteristic M-shape variation in the heat transfer coefficient for two-phase refrigerant flow in microgap channels. Data from [31-33].](/figures/fig-4-characteristic-m-shape-variation-in-the-heat-transfer-1b4s05n0.webp)
![Fig. 13: Average two-phase multiplier (ratio of two-phase to all-liquid heat transfer coefficients) compared with similar data from Rahim, BarCohen, and Ali [27]. Conditions for current study: FC-72, G = 210 kg/m2s, ∆Tsub = 5 K, and a 10 mm by 12 mm by 0.2 mm channel. Conditions for [27]: FC-87, G = 300 kg/m2-s, ∆Tsub = 5 K, and a 10 mm by14 mm by 0.2 mm channel.](/figures/fig-13-average-two-phase-multiplier-ratio-of-two-phase-to-1anx3x05.webp)
![Fig. 1: Improvement achieved by NJTT performers in HEMT RF areal power density, the RF output power over the footprint of the HEMT, as a result of enhanced thermal management enabling operation at higher transistor RF heat fluxes, the heat dissipated over the footprint of the HEMT, adapted from [7]](/figures/fig-1-improvement-achieved-by-njtt-performers-in-hemt-rf-14i5rsk5.webp)
![Fig. 3: Prevailing flow regimes in microscale channels for R134a flowing at 500 kg/m2-s in a 0.79 mm diameter tube: (a) bubbly flow at x = 0, (b) intermittent flow at x = 0.11, and (c) annular flow at x = 0.73 [26]](/figures/fig-3-prevailing-flow-regimes-in-microscale-channels-for-1evy0xeh.webp)
![Fig. 2: ICECool Applications Phase 1 TDV Results [15]](/figures/fig-2-icecool-applications-phase-1-tdv-results-15-jw5vhnz3.webp)