Thermal Management Using PCM-Based Heatsinks
Elon Bauer, Joseph Carlos
Electrical and Computer Engineering
Carnegie Mellon University
Pittsburgh, PA
eob@andrew.cmu.edu, jcarlos@andrew.cmu.edu
Abstract—Existing passive cooling solutions limit the
short-term thermal output of systems, thereby either
limiting instantaneous performance or requiring active
cooling solutions. Active cooling requires relatively large
amounts of power, and cannot easily and efficiently handle
temperature spikes. Phase change materials (PCMs) are
materials that can change phase at a device’s operating
temperature, absorbing energy in the process. PCM-based
heatsinks are able to tolerate short-term spikes in energy
output with substantially less (or even no) active cooling.
In this project we examine the tradeoffs and benefits
present in PCM-based heatsinks by simulating software
workloads and thermal responses on a temperature-aware
simulator.
I. INTRODUCTION
Conventional electronic cooling systems usually consist of
a metal heatsink coupled to a fan, which turns on and off based
on the load in the system. These systems are inefficient at
dissipating heat in a dynamic system that outputs large
amounts of heat for short periods of time. This is due to the
large thermal inertia of the metals used and the delay between
heat input and the time when active cooling begins to take
effect. Phase change materials are materials that are special
only in that they undergo a change in phase at a certain
temperature. During this phase change, PCMs are able to
absorb very large amounts of heat without increasing in
temperature. This ability to absorb energy without requiring
airflow over short periods could lead to a reduced reliance on
active cooling techniques. These materials therefore present a
great opportunity in heatsink design to absorb the heat
dissipated during particular workloads with greater power
efficiency over short time periods than conventional systems.
The major limitations to PCM-based heat sinks are the short
time of phase change and finding a way to contain the PCM in
both solid and liquid states. These limitations will be addressed
and overcome over the course of this paper. Our main focus
will be on the use of Lauric Acid as a PCM, which melts at
44.2 °C, but we will also consider water as a base case. Our
simulation methods include the use of Sniper, McPAT,
HotSpot, and Matlab to arrive at our conclusions. In section II
we will discuss previous work in the field of PCM-based
heatsinks. In section III we will provide a brief technical
description of PCMs and touch on some challenges present in
using them as heatsink materials. In sections IV-VIII we will
discuss our simulation method and in the final sections we will
discuss our results and conclude the paper.
II. PREVIOUS WORK
The use of PCMs for cooling electronics is a new area of
research. Most works focus on creating accurate
parameterizations of various materials or validating their
robustness in real systems, while few have discussed end-toend simulation of the interaction between a chip running a
specific benchmark and the PCM that is cooling it. Tan & Tso
and Kandasamy et al. developed numerical models from
empirical data using different materials (n-eicosane and
paraffin wax, respectively) [1] [2]. Rostamizadeh et al.
developed a theoretical model for PCM from first principles
and validated it with experimental results using calcium
chloride hexahydate [3]. Efforts have been made, specifically
by Wang and Baldea, to simulate power-management systems
using PCMs [4]. However, none of these works involves a
study of how different software workloads might affect the
use of PCM versus conventional cooling methods.
Fig. 1. Temperature of a PCM over time.
III. TECHNICAL DESCRIPTION
Phase change materials are materials which undergo a
change in physical phase when a sufficient amount of heat is
supplied. The specific phase change we will focus on in this
paper will be turning from solid into liquid (melting). During
the time when the material is actively melting, additional heat
applied to the material will not change the temperature of the
material, but is instead absorbed and used in the melting
process. This phenomenon continues until the entirety of the
material is melted. The temperature graph of a PCM over time
looks similar to the graph in figure 1 [4].
Utilizing these materials to absorb heat in their phase
change region is more efficient than using materials which do
not undergo a phase change because during the phase change
an essentially unlimited amount of heat can be input into the
material without the external temperature changing, which is
untrue of typical metallic heat sink materials, which undergo a
linear change in heat as heat is applied.
One of the challenges to this task is that phase change
materials typically also change shape and physical
characteristics as they undergo changes in phase. This poses
challenges for use in mobile systems which have restrictions on
space and do not allow for expansion and contraction of the
materials inside. Existing research has shown that suspending
the PCM inside of a thermally conductive matrix of something
like epoxy provides similar heat sinking capacities to pure
PCM while keeping the overal heatsink from changing shape
[4][3]. This has the added benefit of increasing the surface area
of the PCM, thus allowing it to absorb and release heat more
quickly. We will assume for this paper that this approach is
taken, and that the PCM heatsink can be modeled as a block
consisting of only PCM material.
Another limitation of PCMs is that they can only absorb a
finite amount of heat in their phase change region before they
return to a proportional relationship of output heat vs. applied
heat as shown in figure 1. Thus, PCM-based heat sinks still
need active cooling if the average applied heat over time will
exceed the melting temperature of the PCM. However, if the
application being run consists of short spikes in temperature,
interspersed with periods of low temperature, then the PCM
would be able to disperse the accumulated heat on its own
without needing any active cooling.
Fig. 2. McPAT per-unit power values over time.
VI. HOTSPOT HEAT SIMULATION
McPAT outputs its power values in a JSON format, thus
we were forced to create a Matlab utility to parse the data and
output it in a power trace – “ptrace” – file that HotSpot could
recognize. The code for that utility is included on this project’s
website, listed at the end of this paper.
Beyond that step, most of the process for going from Sniper
to HotSpot is outlined in [7]. We followed that process quite
closely, but with a few exceptions. We created our own
floorplan of the Nehlaem chip, as shown in Figure 3, using the
numbers presented in [7]. However, Sniper did not output as
many functional units as were listed in the paper, so many
things have been lumped into “other” or combined to form
more basic blocks. The Nehalem chip has four cores and an L3
cache, so we concatenated four of our floorplans together and
included an L3 cache, as shown in Figure 4.
IV. SNIPER POWER SIMULATION
In order to simulate the reaction between the PCM-based
heat sink and a CPU for various workloads, we first needed to
obtain power consumption values for a processor under
different workloads. The tool we chose to complete this task
was the Sniper simulator [5]. Sniper has a number of built-in
benchmarks, which can be easily run with different sizes of
data. For our purposes we chose three benchmarks from three
different libraries: Splash2-fft, Parsec-blackscholes, and NPBEP. All three were run with the largest data size, while NPB
was also run with the smaller data size for comparison. We
determined that Sniper by default uses a Nehalem chip, which
we determined would suit our purposes in later steps.
V. MCPAT POWER SIMULATION
By default, Sniper only reports cumulative power values at
the conclusion of the benchmark. In order to force it to produce
values over time, you have to provide the “--viz" and “-power" flags to Sniper. This causes Sniper to invoke McPAT
[6] during its execution to continuously produce power values
over time. In addition to full-chip power values, McPAT
produces per-unit power values for the different functional
units present on the Nehalem chip as shown in Figure 2. This
format turned out to be perfect for use in HotSpot which
produces per-unit heat values by taking in per-unit power
values over time.
Fig. 3. Single core floorplan for Nehalem CPU
From this point we diverge from [7] to compute the heat
numbers generated for the different benchmarks listed in
section IV. With a series of ptrace files generated from Matlab
and a floorplan containing matching blocks, it was easy to run
HotSpot with the basic settings to obtain temperature values
over time. A thermal image of the Nehalem chip running the
Splash2-fft benchmark can be seen in Figure 5. Note that each
core is simply a scaled version of a single core since McPAT
does not produce per-core outputs over time for its per-unit
power values. We ran two iterations for every benchmark, as
suggested in the HotSpot HowTo, to account for initialization
errors as the chip warms up. We also modified the
temperature.c file of the HotSpot code to reflect more accurate
ambient air temperatures and chip startup temperatures.
(K)
Fig. 4. Full-chip floorplan of Nehalem CPU. Dimensions in meters.
VII. PCM ANALYTICAL MODEL
In order to characterize the behavior of a PCM-based heat
sink, we would first need a model that could take temperature
over time as an input and report changes in the heat of the
PCM as well as its phase. We attempted to use the transfer
function derived in [4] to accomplish this goal, however we
eventually determined that the transfer function was not timeinvariant and thus we could not perform a Fourier Transform
on it to perform calculations in the frequency domain.
After this setback, we came to the realization that we would
need to create an analytical model from scratch by ourselves.
We did so using Matlab. The model assumes that the heatsink
is a 7 cm by 7 cm by 3.5 cm box with a copper container. This
container is attached to the chip via thermal paste, which is
assumed to have a significantly higher thermal conductivity
than copper.
𝑑𝑄
𝑇! − 𝑇!
𝐻=
=𝑘∗𝐴∗
(1)
𝑑𝑡
𝐿
Equation 1 gives the rate of transfer of energy through a
surface of length L and area A given the "hot" region's
temperature, TH, and the "cold" region's temperature, TC.
Since the simulation results from HotSpot are temperatures
sampled at a static sampling interval, we can solve for the
energy input into the heatsink from the chip at a given
timestep, as shown in equation 2.
Fig. 5. HotSpot heatmap of Nehalem CPU running Splash2-fft.
𝑑𝑄 = 𝑑𝑡 ∗ 𝑘 ∗ 𝐴 ∗
𝑇!"# − 𝑇!"#
(2)
𝐿
In equation 2, k is the thermal conductivity of the material
separating the CPU and PCM (in our case, copper), A is the
area of the chip connected to the PCM, which is just the chip's
total area, and L is the width of the wall of the copper PCM
container.
A similar quantity can be obtained for the heat transfer due to
the ambient air. Adding these quantities results in the total
energy transfer in the system in a single time step.
𝑄 = 𝑚 ∗ 𝑐 ∗ 𝑑𝑇 (3)
Equation 3 represents the heat required (Q) for a temperature
change in a mass m. Here, c is the specific heat of the material
and dT is the change in temperature. Using the specific heat of
the PCM and the mass of the PCM (calculated using the
volume of PCM and it's density), this equation can be
rearranged to give equation 4, which is the change in
temperature of the PCM in a single time step. Q in this
equation is found using equation 2.
𝑑𝑇 =
𝑄
(4)
𝑚∗𝑐
When the PCM is not undergoing a phase change, the change
in temperature can be obtained using equation 4. During this
phase change we use equation 5 to determine whether the
entire PCM has melted or solidified. m is the mass of the PCM
and L is its latent heat of fusion. During a phase change, the
PCM has constant temperature, so we do not use dQ (from
equation 2) to find dT. Instead, we accumulate dQ until the
necessary amount of energy (m*L) has been accumulated or
dispensed by the PCM. At this point, the PCM has changed
phase, and we resume using equations 2 and 4 to find the
change in temperature.
𝑄 = 𝑚 ∗ 𝐿 (5)
else if state is liquid:
dT = dQ / (m * c_liquid)
if T_PCM + dT < T_melt:
T_PCM_next = T_melt
state_next = changing phase
else if state is changing phase:
Q_acc_next = Q_acc + dQ
if Q_acc + dQ > m * L_PCM:
state_next = liquid
Q_acc_next = 0
The pseudocode for our model’s state machine is as follows:
Inputs:
CPU_temp_input = list of input CPU temperatures, one per
sampling interval
dt = sampling interval
ACPU = area of CPU
Aair = area of PCM in contact with air
Lbox = width of copper container wall
k = thermal conductivity of copper
Tair = constant
m = mass of PCM
cliquid = specific heat of PCM in liquid phase
csolid = specific heat of PCM in solid phase
LPCM = latent heat of fusion of PCM
TPCM = initial PCM temperature
state = initial PCM state
Outputs:
PCM_temp_output = list of output PCM temperatures, one
per sampling interval
Pseudo code:
Qacc = 0
while CPU_temp_input is nonempty:
TCPU = pop(CPU_temp_input)
TPCM_next = TPCM
state_next = state
Qacc_next = Qacc
dQCPU = dt * k * ACPU * (TCPU - TPCM) / L_box
dQ_air = dt * k * A_air * (T_air - T_PCM) / L_box
dQ = dQ_CPU + dQ_air
if state is solid:
dT = dQ / (m * c_solid)
if T_PCM + dT > T_melt:
T_PCM_next = T_melt
state_next = changing phase
else if Q_acc + dQ < - m * L_PCM:
state_next = solid
Q_acc_next = 0
push(PCM_temp_output, T_PCM)
T_PCM = T_PCM_next
state = state_next
Q_acc = Q_acc_next
The heat transfer equations in this section were found in [8].
VIII. PCM SIMULATION
The final step towards simulating our PCM-based heatsink
was to find a PCM with thermal properties that would perform
well as a heatsink at the temperatures the chip was
experiencing. A list of PCMs and their thermal properties can
be found in [9], and based on the fact that for all benchmarks
the CPU was running at around 50 °C, we determined that
Lauric Acid, with a melting point of 44.2 °C would be an ideal
PCM for our purposes. We also include water as a baseline.
As mentioned in section VII, the PCM analytical model
uses the heat of fusion, the specific heats of solid and liquid
phases, the melting point and the density of the material to
characterize its response to temperature input. This provides
an accurate response characterization and shows how the
phase change would occur in the presence of heat from the
CPU and cold from the ambient air around it.
IX. RESULTS
In order to get a wide range of results, we created
temperature profiles of the Nehalem chip in HotSpot for four
different benchmarks. These were Splash2-fft with the large
dataset, Parsec-blackscholes with the large dataset, and
NPB3.3-MPI (EP) with both the large and small datasets. The
differences in chip utilization between these benchmarks led
to interesting variations in the performance of the PCM as a
heatsink. However, because HotSpot produces very small
variations in temperature, we increased the variation in
HotSpot’s outputs. We also set the room temperature for our
PCM to be very close to the melting point, simply because this
makes it easier to observe the melting reaction in graph form.
We found that for benchmarks like Splash2-fft, which
exhibits CPU temperatures as shown in Figure 6, the PCM
behaved in an ideal manner. The PCM response for Splash2fft is shown in Figure 7 and shows that the PCM reaches its
melting point very rapidly, but then stays at a constant
temperature due to the fluctuations of the CPU temperature.
With a workload that fluctuates such as this, the PCM will
never gain enough heat to completely melt and will thus never
increase in temperature. This is the ideal scenario.
Fig. 8. Parsec-blackscholes CPU temperature response
Fig. 6. Splash2-fft CPU temperature response
Fig. 9. PCM heatsink response to Parsec-blackscholes benchmark
is for many iterations of the Parsec benchmark, which only
runs for a few milliseconds for a single run. So, if the load is
high, but for a short time, i.e. a temperature spike, PCM-based
heat sinks should be able to effectively dissipate that heat.
Fig. 7. PCM heatsink response to Splash2-fft benchmark
A less ideal scenario is one like that shown by the Parsec
benchmark which climbs quickly to a high load and maintains
that load for a long time. This causes the CPU to stay at a high
temperature (about 51 °C in this case) for a very long time as
shown in Figure 8. Because this temperature is greater than
the melting temperature of the PCM, the PCM completely
melts and then resumes a linear relationship between heat
input and output, eventually reaching equilibrium with the
ambient air around 44.3 °C after 120 seconds of simulation
time as shown in Figure 9. This is the worst-case scenario,
since the PCM is not doing any better than a regular heatsink
after it completely melts. However, it should be noted that this
The NPB benchmarks that were run behaved similarly to
the Parsec benchmarks. If only run for 5-10 iterations of the
benchmark, the PCM stayed in the phase transition and
perfectly absorbed all the heat. However if the benchmark
were run for long enough, it too would completely melt the
PCM and cause it to enter the linear region.
As a reference, and to verify our model, we created a
synthetic temperature profile for use with water as the PCM.
The temperature profile starts below the melting point of
water and then jumps to 40 °C so that the water can
completely melt, then it goes back down so we can see the
water freeze again. The temperature profile of the chip and
PCM are shown in Figures 10 and 11 respectively. These
figures demonstrate that our analytical model is sound and
performs as expected, thus validating the rest of our results.
X. CONCLUSION
In this paper the capability of phase change materials to
function as heatsinks for CPUs under various workloads has
been explored in detail. Power utilization values were
obtained using the Sniper simulator and McPAT for various
benchmarks running on a Nehalem CPU. A floorplan of that
CPU was created and used for HotSpot simulations, which
output temperature values of the various units on the chip.
Finally these temperature values were inputted into a custombuilt analytical model of a Lauric Acid-based heatsink to
characterize its performance as a heat dissipater. The model
was validated with water being used as the phase change
material and a synthetic heat input.
will become proportional to input temperature. This is not a
huge problem for short periods of time, since in this region the
PCM is simply performing similarly to how a regular heatsink
would perform.
The field of PCM-based heatsinks is a growing and
exciting field and the authors hope that much more research
will be done towards making these thermal wonders a reality.
XI. MILESTONES AND SCHEDULE
Table 1 contains a list of project milestones and the dates
by which they were completed. The year is 2014.
TABLE I.
PROJECT MILESTONES
Milestone
Obtain or create a thermal model of a processor for use
in HotSpot and Sniper.
Obtain power numbers for various benchmarks through
Sniper and McPAT.
Create a floorplan of the processor for use in HotSpot
Obtain heat values over time from the processor using
HotSpot.
Obtain or create a simulation model for a PCM-based
heat sink.
Obtain hard numbers for PCM-based heat sink
performance.
Analyze PCM-based heatsink performance and
comment on feasibility
Final paper.
Date
November 11
November
11/20
December 3
December 3
December 6
December 7
December 9
December 11
XII. WORK DISTRIBUTION & WEBSITE
Fig. 10. Synthetic temperature response for use with water
The work was split evenly between the authors for the
entirety of the project. Joseph researched and discovered how
to make Sniper output power numbers over time, and handled
everything related to Matlab including the script for parsing
McPAT output and the PCM analytical model. He was also
the primary researcher, contributing most of the previous work
section and finding out what the Nehalem CPU specifications
were. Elon found benchmarks suitable for the application and
ran the simulations for Sniper, converted and input that data
into HotSpot, and created Python scripts to make the output
consistent with the PCM model inputs. He also created the
floorplan for the Nehalem CPU for use in HotSpot. Both
remembers contributed to writing the report and creating the
website.
Our project website was created by Joseph and Elon and
can be found at http://www.andrew.cmu.edu/user/eob/. We
hope you enjoy it.
Fig. 11. Water-based heatsink melting and freezing response
We determined that properly selected phase change
materials can be highly effective at dissipating heat generated
from a CPU over relatively short periods of time. If there are
short spikes of heat interspersed with temperatures well below
the melting point of the PCM, the PCM will stay in the phase
change region and will stay at a constant temperature. If,
however, the input temperature is consistently very high, the
PCM will melt completely and the response to temperature
REFERENCES
[1]
[2]
[3]
[4]
F.L. Tan, C.P. Tso. “Cooling of mobile electronic devices using phase
change materials,” Applied Thermal Engineering vol. 24, issues 2-3, pp.
159-169, February 2004.
Ravi Kandasamy, Xiang-Qi Wang, Arun S. Mujumdar. “Application of
phase change materials in thermal management of electronics,” Applied
Thermal Engineering, vol. 24, issues 17-18, pp.2822-2832, December
2007.
Mohammad Rostamizadeh, Mehrdad Khanlarkhani, S. Mojtaba
Sadrameli. “Simulation of energy storage system with phase change
material (PCM),” Energy and Buildings, vol.49, pp.419-422, June 2012.
Siyun Wang, M. Baldea. “Storage-enhanced thermal management for
mobile devices,” American Control Conference (ACC), pp.5344-5349,
June 2013.
[5]
[6]
[7]
[8]
[9]
Sniper Power Simulator, http://snipersim.org/
McPAT Multicore Power, Area and Timing Framework,
http://www.hpl.hp.com/research/mcpat/
Adrian Florea, Claudiu Buduleci, Radu Chis, Arpad Gellert, Lucian
Vintan. “Enhancing the Sniper Simulator with Thermal Measurement,”
18th International Conference on System Theory, Control and
Computing, Sinaia, Romania, Oct. 17-19, 2014
Hugh Young, Roger Freedman, Lewis Ford. University Physics with
Modern Physics (12th Edition). Benjamin Cummings, 2008.
Wikipedia, “Phase-change Material, Thermophysical Properties of
Common PCMs,” http://en.wikipedia.org/wiki/Phase-change_material.