CED Simulations of Heat Transfer from a Heated Modul
Experiments and a Parametric Study
Masud Bebnia’, Wataru Nakayama”, and Jeffrey Wang”
* Schoo! of Mechanical and Manufacturing Engineering
‘The University of New South Wales, Sydney 2052, Australia
an Air Stream: Comparison with
‘** CALCE EPRC, University of Maryland, College Park, MD 20742,USA
Abstract - In the last decade or so CFD simulations have
‘become more and more widely used in studies of electronic
cooling. Validation of these simulations has been considered
to be very important. Nakayama and Park have performed a
series of experiments with the aim of measuring the heat
‘wansfer and flow field in a simple geometry of a heated chip
in anair steam, They observed a number of complex features
in both the flow field and heat transfer behavior. In this study,
their experimental geometry has been used to simulate the
flow and temperature fields in a parallel-plate channel with a
Iheated block mounted on the floor. The channel inlet flow
velocity has been varied between 1 and 7m/s. Various
turbulence models have been tested, and the effect of the
channel inlet flow on the heat transfer rate has been
‘determined by considering both a uniform and fully-developed
condition, The substrate adiabatic heat transfer coefficient is
also numerically determined. The results indicate that the
flow in the vicinity of the module is three-dimensional, and
exhibits flow separation and vortex formation, hence leading
to.a complex distribution of the local heat transfer coefficient
fon the substrate, The air temperature next 10 the floor is
strongly affected by the heat transfer from the block, which
leads to the formation of a thermal wake downstream of it,
‘The experimental dala is used to validate the CFD predictions
ind the agreement for some parameters is shown to be
INTRODUCTION
‘The evolution of the electronics has led to a growth of this
industry and utilization of microprocessors in numerous
applicaions. This in tum has made the microelectronic
packaging an extremely important technology with a very
Tapid Jace of evolution and diversification. The market
demands of the information age are such that the cycle of
product must become ever-increasingly shorter.
‘To meet such stringent demands, highly efficient tools are
required forthe design and manufacturing. To this end, a host
of techniques, design tools and software have been developed
and are now indispensable tools for meeting the marketplace
‘demands. ‘These also need to be continuously improved and
upgraded which leads to computers of today breeding
Computers and other electronic devices of the next generation.
It is perhaps fair to say that, the development and application
of software tools in the electrical aspects of packaging has
‘been more advanced than those dealing with cooling
problems. Those programs to deal with cooling problems
‘made relatively recent arivals inthe industry design room.
With the advent of Computational Fluid Dynamics (CFD) in
the recent years, flow and heat tansfer computations have
0-703-447589810 0001908 EEE 143
become quite readily possible. In particular, with the recent
‘nuoduction of high power workstations and personal
computers the cost of such computations has been drastically
reduced and as a result many CFD codes have come into the
market. More recently, such computations have become very
popular in the application area of cooling of electronic
‘components. In fact, this popularity has led to several purpose
‘developed CFD codes coming into the market which are
specifically tailored for use by heat transfer engineers in the
electronic industry.
‘The heat dissipation from gatearray chips continuously and
sharply increased in the 1980's and this demanded more
efficient and highly optimized cooling techniques. Although,
this chip power tend halted in the 1990's (1), the circuit
integration and size reduction of the systems reinforced the
importance of thermal management and development of tools
for this purpose (2]. From a modeling point of view, the
progress of circuit integration technology as worked to
simplify the geometric situation on printed circuits. In the
ultimate situation all power-consuming functions are
concentrated in a single module. To some extent, this has
already become a reality in certain small systems, where a
‘micro-processor chip is a single intense heat source on a
‘board, and it is surrounded by memory modules which
dissipate far less heat than the former. This simplification
does not necessarily make the problem solution any simpler,
a a concentrated heat source leads to an increase in reliance
‘on the heat spreading function ofthe substrate which requires
‘a conductive-convective conjugate analysis technique and this
{ntum increases the numberof design parameters and solution
time.
‘The problem of cooling of heated modules has been widely
studied. In particular, in the last few years a number of
researchers have used CFD w study various aspects of the
flow and heat transfer in this geometry. An extensive survey
of literature is not intended here, however some of the more
recent studies [3,4 and 5] refer t0 previous relevant research.
Some researchers have also compared their numerical
simulations with experiments. One of the recent studies inthis
category is by Anderson [6] who used a commercial CFD
code to simulate a row of heated modules. Although, a
reasonable agreement forthe adiabatic heat transfer coefficient
Of the module was obtained, the supsrposition kemel function
(hich is a measure of how hot one module becomes due 10
heating of another module) was as much as 50% higher than
the experimental results. This was attributed to the fact that
the code did not appropriately resolve the amount of cross
channel mixing. The author notes thot this aspect needs to be
1998 interSoiety Conference oa Thermal Phesomestfurther studied and that obtaining accurate results for realistic
applications may require a prohibitive number of grid points,
‘The relatively recent adoption of CFD simulation studies in
electronic cooling applications has prompted some interes in
validation of these codes for these types of problems. In
‘general, validation and benchmarking of CFD codes has been
‘an on-going research area attracting a Jot of attention from
both users and code developers. To this end, a number of
classical problems have been adopted and purpose-developed
experimental databases have been constructed. for this
purpose. One of the more recent reports on the validation of
commercial CFD codes is presented by Freitas [7].
Benchmarking and validation for electronic cooling
applications is also of great importance because of the
‘complexity of physical phenomena. This has been recognized
as very important as evidenced by a number of recent
publications (e.g. ASME HTD-Vol. 255 [8] and ASME HTD-
Vol. 292 [9}) and panel discussions in conferences (e.g.
Iherm’96, ASME-WAM'96 and ASME National Heat
‘Transfer Conference’97). One of these cases is proposed by
Nakayama and Park [10] who studied a block (simulated
module) mounted on the floor (substrate) of a rectangular
channel, They developed a novel synthesized approach by
considering the conjugate heat tansfer in two module/floor
combinations, ie. a heated module on an adiabatic floor and
‘an adiabatic modale on a heated floor. ‘They also pesfoomed
detailed measurements of velocity and temperature of air and
cenain key features of their data set can be employed to test
CED codes used for thermal design analysis. The details of
heir experimental measurements and analysis can be found
‘elsewhere {10 and 11].
In our study, we have considered the problem of air cooting of
‘a heated module located on the floor of an adiabatic
rectangular channel. “The geometry is similar to that of
Nakayama and Park [10] s0 that comparison with their
experimental measurements can be made, ‘Three-dimensional
turbulent flow simulations are performed for a range of
channel inlet velocities, Particular attention has been focussed
fon the local flow and heat transfer behavior in the vicinity of
the module. The local heat tansfer coefficient on the module
land the local adiabatic heat transfer cooffcient on the floor
hhave been computed. ‘The effects ofthe channel geometry and
inlet condition on the heat transfer and flow have been
‘quantified. The results are useful in determining the merits
and shortcomings of CFD simulations. Although, a conjugate
Conduction-convection analysis has not been considered here,
an extension of this stdy to property account for conjugate
effects in the future is essential,
1. THE PROBLEM
‘Nakyama and Park performed their experiments using a Block
(Gimulated module) mounted on the floor (substrate) of &
parallel-plate channel. Heat was genorated uniformly over the
Dottom surface of the block at the sate Qu and a part of the
‘heat generation was transferred from the Block's surface to the
air flow over the block, called direct heat transfer and denoted
8 Q,. The rest, Q, {Gund its way through the block support,
‘the channel floor, then from the floor surface to the airflow in
144
the environment. ‘The bottom of the channel was made
adiabatic, 50 that all Q, ended up inthe ar low in te channel.
In prctical situations. seen in electonic equipment Q,
‘amounts 10 a substantial fraction of the total heat generation
Qu. A dicct estimation of Q, is challenging ask. Fist of
alll the flow near the module is thrs-dimensional and
involves flow separation and vonex fomation, hence the Tocal
heat transfer coefficient onthe floor has a complex
distribution. Besides, the air wmperatue next 10 tho floor
Surface is affected by the reas of Q, fom the block,
taney, the covering ofthe floor by thermal wake originated
from the block. ‘The two heat flow components, Q, and Q,,
a coupled, and the partion of Q, into Q, and Q, is
function ofthe block dimension, the air velocity the them
resistance of the block support, the thickness ‘and thermal
conductivity of the floor. The block support is a mode ofthe
clectrcal interconnects between the electronic module and the
Wiring subswate, henes, its thermal conductance varies with
the numberof interconiecis and the material and dimensions
ofthe eleecal leads. A high thermal conductance can be
realized where adhesive or thermal paste is insemed at the
‘odule/ubsits interface in order 10 enhance heat transfer
through the substrate, Meanwhile, te wiring substrate may
have a varity of thermal conductance depending on the
‘ising capacity and the materials of electrical conductors and
insulation matrix, Therefore, possible combinations of
relevant parameters amount to avery large numer.
‘The computing time and memory capacity needed for the
solution of a conjugate heat transfer problem are far greater
than those for the solution of a non-conjugale counterpart.
‘Nakayama and Park proposed a scheme by which the demand
on the computational or experimental resource can be reduced
drastically. Their method of heat transfer estimation uses two
basic quantities, the themnal wake function, F,, and the
adiabatic heat transfer coefficient, hy, They determined these
parameters experimentally using two combinations of block
‘and floor, a heated block on an adiabatic floor (HB/AF) and an
‘adiabatic’ block on a heated floor (AB/HF). The HB/AF
experiment produced the information about F-. They used a
copper block with a square footprint area of 31x 31 mm’, and
a height of 7 mm which was located on the centerline of a
parallel-plate channel whose cross section was 320 mm wide
nd20 mm high. Their wind tunnel was a suction tne, with
a bellsmouth atthe entrance located 120mm upstream of the
block. They performed detailed turbulent velocity
measurements by a hot-wire snemometer and liquid crystal
thermography produced the information on the local adiabatic
floor temperature and heat transfer coefficient.
‘They took maximum care in the experiments to impose the
necessary thermal boundary conditions as accurately as
Possible. However, certain compromises and assumptions
‘were inevitable to ease the conduct of experiment. The basic
assumption was that h, can be approximately detemnined on
the uniform q, floor. “Namely, the thermal wake originated
from the upstream area of the floor has negligible influence on.
hh on the downstream area instead the thermal wake from the
block prevails. Meanwhile, a rigorous definition of h needs
situation where only a spot area of the floor is heated, and all,
the rest is made adiabatic. To determine the distribution of hy
1098 tterSocity Conference on Thera Phenomenaofthis definition, the heater spot has 1o be moved around over
the floor area. This is a prohibitively demanding task, if one
attempts to find a distribution experimentally. Besides. the
Theat leakage from the heater to the surrounding thermal
insulation introduces greater uncertainty asthe heater is made
smaller. Considering these experimental difficulties and
‘uncertainties, a recourse to CFD simulations for determining
the effect of these parameters seems inevitable. Also, the
CED simulation can be used to justify the assumptions
involved in the experimental set-up. The heated-block on an
adiabatic floor configuration was simulated here.
IIL THE GEOMETRY AND CFD MODELING
In dhe simulation of many experiment, it i recognized that
cone can not possibly perform the computations in the same
domain as the experiments were performed because of the
fxtent of tbe experimental domain. Therefore, @ sound
jdgment is required for the selection of the computational
domain such that the important flow and heat transfer
characteristics revealed bythe experiments are nt affected by
the computational boundaries. Once a. reasonable
computational domain is selected it needs to be tested by
changing the dimensions to ascenain its adequacy. For the
above experiment, with the relatively large physical dimension
of the duct compared to the chp it was necessary 10 choose &
reduced size computational domain of 271x131x20 mm. In
this geometry, the module was equally distanced from the inlet
and outlet ofthe channel (Le. 120mm each), ‘The domain size
in the x and y directions were doubled and it was noted that
this made very litle difference on the ow and thermal fields.
The block was an isothermal body of 31%31x7 mn at 353K
placed on the adiabatic channel floor and the cooling air
Entered the channel at 293K. The symmetry plane through the
channel centerline was employed to reduce the computational
effor.. The geometry and computational domain is showa in
Fig. 1 Gin tis igure the distance berween the inlet and the
chip is reduced to 29mm as will be explained later).
For the CFD simulation we used the commercial package
FLUENT [12] which is based on a boundary fitted Gite
volume technique. It is @ general-purpose code with an in-
built mesh generation capability for rectangular geometries.
‘There are two versions of this package available, structred
and unstructured grid. We used the structured gyid version to
‘Generate the grid and obtain preliminary solutions. Then, the
{rd and solution was imported into the unstructured version to
focally refine the gid for a more accurate resolution of high
velocity and temperature gradients near the heated block.
‘There are thee turbulence model options (ie. ke, RNG and
Reynolds stress). The primitive-varables form of the Navier-
Stokes or Reynolds-averaged Navier-Stokes equations are
integrated in each control volume adopting the divergence
theorem. For accuracy, we used the QUICK diseretizaton of
the convective terms (eat least second order accurate).
In turbulent flow calculations, the wall boundary conditions
are extremely important for the accurate prediction of heat
transfer. The heat transfer rate in a turbulent flow is strongly
affected by the wall treatment (e.g. see Ciofalo and Collins
[13] and Djilali etal (14). FLUENT has the option of either
14
Fig. 1. The geometry and computational domain.
using a wall function ora two-layer zonal model which allows
integration to the wall. In general, for the computations with
the empirically-derived wall functions, the frst grid point
should be placed outside the viscous sublayer (ie. y>I1).
‘This may not be feasible in some flows as the ply
location ofthis grid point may be well into the computational
domain. Further, the integration all the way to the wall (Le.
‘two-layer model) imposes a severe restriction onthe near-wall
arid spacing by requlring several grid points in the viscous
sublayer witha careful placement of the nodes. ‘The first grid
pint is to be placed such that y" is ofthe order of unity at this
ocation [12], We ensured that this was the case in our
‘computations.
AIL the simulations presented here were performed on an
HP9000 workstation. They are all three-dimensional
‘computations using non-uniform meshes with the grid lines
‘being appropriately biased 10 resolve the sharp gradient
regions (@. near the solid boundaries). In order to obtain
accurate solutions, local grid modification was performed by
using a hanging-mode refinement. The area around the
‘module was in particular important for the resolution of the
local gradients, Tt was insured thatthe restriction on y" noted
above was satisfied, in addition to obtaining no more than a
(0.5% change in the average Nusselt number on the module
when the grid was refined. As a result of this mesh sensitivity
analysis we ended up using up 10 180,000 elements (the
surface grid on the module is shown in Fig. 2). This is indeed
a fine gid and ensures an accurate resolution of the local
velocity and temperature gradients. We performed some
‘preliminary computations and compared the ke and RNG
turbulence models, There was not a significant difference
Detween the two models and we chose the RNG model as it
performs somewhat better for recirculating flows (12). As
noted previously, for this flow, wall-functions are not suitable
(this has also been noted by Anderson [6]) and the two-layer
‘model was adopted which meant a very fine grid requirement
around the module and on the channel floor. For all of our
simulations reported here, we made sure that this requirement
was satisfied.
IV. RESULTS.
In our simulations we varied the inlet air velocity in
accordance with the experiments in the range of 1 wo Tm/s.
10098 terSociety Conerence on Thermal PhenomenaFig. 2. The surface grid on the module,
One of the difficulties encountered in these types of
simulations for comparison with experiments isthe imposition
ff the correct boundary conditions. This is in particular
imponant when the geometry of the experimental setup can
not be exactly replicated, In most studies itis customary to
use a uniform inlet velocity (i.e. a plug profile). In the
experiments as noted above, a bell-mouth was used and due to
the boundary layer development on this surface, the inlet
velocity profile would not be a uniform one. In order to
examine the effect of this inlet velocity condition we
performed two types of computations: @) uniform inlet
‘velocity and (ii) fully-developed inlet velocity. For both of
‘these, te inlet turbulence level was set to the measared value
(ie. 10%). For the second type, the exact computational
domain was as given in Fig. 1. ‘In the second series, we
‘biained a fully-developed velocity profile by using the same
cross-section channel but without te module, In this channel,
we perforned a number of repetitive computations by using
‘the outlet velocity as the inlet tla fully-developed profile was
achieved. Then, this profile was used as the inlet to a channel
‘with the module. Since in these simulations the inlet profile
\was fully-developed the inlet section length was shortened 10
29mm (instead of 120mm) in order to save on computational
fost. The effect of the duct exit was also examined by using
fn outlet boundary condition and a pressure. boundary
condition. The effect for the domain considered here on the
flow and heat transfer was marginal and we adopted the outlet
condition forthe computations performed.
Fig. 3 shows the velocity vectors on the symmetry plane for
the uniform and fully-developed inlet velocity profiles. In
‘both cases, the recirculation zone behind the block is noted.
FFor the uniform inlet velocity the length of this zone was
slightly longer. In the front of the module, in both cases a
‘small recirculating zone was observed, with the uniform inlet
146
fully developed flow inet
Fig. 3. The computed velocity vectors on the symmetry
plane for an inlet velocity of Tin,
») fully detveloped flow inlet
Fig. 4. The computed velocity vectors on a plane
parallel to the channel floor at Z2=3mm for
aninlet velocity of im/s.
indicating a larger zone with higher velocities. On top of the
‘module, a its leading edge a small separation zone was noted
with the uniform inlet exhibiting a larger one. “This is more
Clearly evident in Fig. 4 showing the velocity vectors on a
plane parallel to the flow channel floor half-way through the
‘block. The horseshoe vortex downstream of the module
‘observed in the experiments is clearly evident.
“The computed axial velocity profiles at two y locations of 50,
‘and 65.Smm for the uniform inlet condition of 3mys is
‘compared with the experimental measurements in Fig. §. The
two locations corespond to the side of the chip and the
centerline, respectively. It is noted that there is a reasonable
‘agreement between the predictions and measurements. The
flow recirculation right upsream of the module is evident in
the simulated results whereas it is not evident in the
‘measurements. Further, right downstream of the module the
flow reversal predicted in the simulation is not clearly
measured by the hot wire anemometer. This is perhaps
1996 nrSacety Conference on Therma PheomesaVelocity
ofthe chp). V 23 (ove)
x Com
Velocity
‘Y= 85.5 mm (conrne ofthe chip), V=3 (ra)
4 3
co Tea toe 0
cm
Fig. 5. The velocity profil at two y locations fora uniform inlet velocity of 3 mis.
Velocity
Y= 80 mm (de ofthe chip), Vn (rv)
ae 70 78
Velocity
mm (Centarine of he ep). Vn 3 (vad
im)
(mm)
(amy
Fig. 6. The velocity of profiles at two y locations fora flly developed inlet velocity (3m).
attibuted to the accuracy of the measurements in this region
of low and reversed velocity. The velocity profile recovery