Fingerprinting For Chiplet Architectures Using Power Distribution
Fingerprinting For Chiplet Architectures Using Power Distribution
Fingerprinting For Chiplet Architectures Using Power Distribution
ScholarWorks@UMass Amherst
August 2023
Part of the Signal Processing Commons, and the VLSI and Circuits, Embedded and Hardware Systems
Commons
Recommended Citation
Burke, Matthew G., "Fingerprinting for Chiplet Architectures Using Power Distribution Network Transients"
(2023). Masters Theses. 1330.
https://doi.org/10.7275/35660113 https://scholarworks.umass.edu/masters_theses_2/1330
This Open Access Thesis is brought to you for free and open access by the Dissertations and Theses at
ScholarWorks@UMass Amherst. It has been accepted for inclusion in Masters Theses by an authorized
administrator of ScholarWorks@UMass Amherst. For more information, please contact
scholarworks@library.umass.edu.
FINGERPRINTING FOR CHIPLET ARCHITECTURES
USING POWER DISTRIBUTION NETWORK
TRANSIENTS
A Thesis Presented
by
MATTHEW BURKE
May 2023
A Thesis Presented
by
MATTHEW BURKE
I would first like to thank the MITRE Corporation for funding this project, and
for their oversight and assistance. Second, I want to thank Wayne Burleson for his
long-time role as my advisor and mentor, for pushing me past my limits and for
providing me the opportunity to become an electrical engineer. Additionally, I thank
Dan Hoclomb for the long-term guidance he has provided me on this project, and
my colleagues for their companionship, assistance and inspiration for this work. In
particular, I thank Max Cohen Hoffing for his supporting experiments with hardware
implementations of our work. I thank my parents and the rest of my family for their
life-long support and provision. Everyone here made this work possible.
Above all, I thank the Triune God; with Him I have everything, without Him,
nothing.
iii
ABSTRACT
MAY 2023
MATTHEW BURKE
B.Sc., UNIVERSITY OF MASSACHUSETTS AMHERST
several small, interacting chips on an interposer rather than the traditional, single
chip used for a device. Like any other type of integrated circuit, chiplets are in need
of a physical layer of security to defend against hardware Trojans, counterfeiting,
transient signals within the PDN to identify individual chiplet systems and physical-
layer threats against these devices.
In this work, we describe a Python-wrapped HSPICE model we have built to
iv
to identify unique characteristics in the voltage response signals to transient stimuli.
We provide the true positive and false positive rates of these methods for a simu-
lated lineup of chips across varying operating conditions to determine uniqueness and
reliability of our techniques.
Our simulations show that, if characterized at varying supply voltage and tem-
perature conditions in the factory, and the sensors used for identification meet the
sample rates and voltage resolutions used in our tests, our protocol provides sufficient
v
TABLE OF CONTENTS
Page
ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
CHAPTER
1. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 PUFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 PDNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Chiplets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Our Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Structure of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2. RELATED WORK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Transient Signal Sensing/Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Chiplet PUFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
vi
4. SIMULATION RESULTS AND SIGNAL PROCESSING . . . . . . . . . 19
4.8.1 AC Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5. HARDWARE EXPERIMENTATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7. CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
vii
LIST OF TABLES
Table Page
3.1 Table of parameters used for our PDN models, taken from [1]. The
values correspond to those in figure 3.4 * nominal capacitance
value is reduced by a factor of 2 to mimic active decap being
disabled. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.1 True positive rate/false positive rate of the time domain L1 norm
identification method, given as a function of sample rate and
voltage measurement precision. Rates given as cumulative across
temperature and supply voltage corners. 1.0 indicates a 100%
rate, and 0.0 a 0% rate. *: indicates points where fingerprinting
seems to fail. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 True positive rate/false positive rate for time domain L2 norm. Same
format and method as 4.1. *: indicates points where
fingerprinting seems to fail. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.4 True positive rate/false positive rate for CWT L2 norm. Same format
and method as 4.3. *: indicates points where fingerprinting seems
to fail. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
viii
LIST OF FIGURES
Figure Page
3.1 Diagram of the top-layer grid for the single-die model, with the
interposer connected to its edges. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Diagram of the bottom-layer grid for the single-die model, with
NMOS sinks. The data here normally assumes these sinks to be
in the center. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Unit cell model of metal layers in a PDN, with transmission lines
used for wires [4]. Note: this particular model does not include a
conductive component (G) in parallel with the capacitor. . . . . . . . . . . 14
3.4 Diagram of a unit cell, with the ground wire RL values lumped into
those of the the power wires [1]. With explicit power and ground
wires, each would have nominal resistor and inductor values of R
and L. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.5 Diagram of the RLGC model for a through-silicon via (TSV) used in
[5]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.1 Diagram of the dice (sink, suspect, trusted) and interposer (grey)
arrangement used for our simulations [6]. Sinks are placed across
each point of the sink die, and measurements are made on the
edge near the suspect (blue). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
ix
4.3 Zoomed-in view of the response to the current sinks’ rising edge,
shown in figure 4.2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.4 Zoomed-in view of the response to the current sinks’ falling edge,
shown in figure 4.2. Notice the variation becomes more noticeable
here; this is where we see most variation in our wavelet
transforms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.5 PDN response to the rising edge of the sink, plotted at various
temperature and supply voltage conditions. Gathered from the
single-die model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
x
4.13 Scatter plot of the resonance peaks shown in figure 4.12. . . . . . . . . . . . . . . 36
4.15 Scenario with trusted dice (sink and sensor) on opposite sides of a
potentially hostile die (suspect) [6]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.1 The Virtex UltraScale+ FPGA used for the supporting hardware
implementation [7]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.1 Lumped model of an IC’s (not chiplet) PDN, with workload [8]. . . . . . . . 43
6.2 Sketch of the droops from the circuit modeled in 6.1 in the time and
frequency domain [8]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
xi
CHAPTER 1
INTRODUCTION
1.1 PUFs
Physical unclonable functions (PUFs) are hardware-based systems intended to
identify an individual chip apart from any other, and prevent counterfeiting and/or
identify when a chip has been tampered with. Given a physical stimulus (chal-
lenge), they produce a measurable physical response that- if designed and imple-
mented correctly- should be unpredictable to an attacker. Weak PUFs and strong
PUFs are the two broadest classes of PUFs; the former rely upon a single (or
few) challenge-response pair(s) (CRPs), while the latter rely upon a massive set of
challenge-response pairs- ideally ones that exist on an exponential space and contain
many more CRPs than can be exhaustively tested by an attacker [9]. The response-
or series of responses- should be unique to each chip, and any tampering should
change the PUF’s response sufficiently to classify it as hostile. However, it must also
be reliable- that is, produce sufficiently similar responses across a wide variety of
1.2 PDNs
Power Delivery Networks (PDNs) are critical to the function of any modern inte-
grated circuit. As their name implies, they are responsible for distributing sufficient
power for the chip’s operations, as well as regulating power supplies to provide con-
sistent power supply and protect the chip from damage [10]. PDNs are typically
extremely complex, multi-layer grids of wires and nonlinear components to ensure
1
Figure 1.1: Lumped Diagram of an FPGA PDN [2].
stable power is delivered to the entire chip. Since these structures are ubiquitous and
integral to the entire chip, they make promising candidates for PUFs (or other fin-
gerprinting methods); tampering with a chip would practically necessitate disturbing
its PDN, and using them as a PUF would require little to no additional hardware or
logic.
1.3 Chiplets
Due to the increasing difficulty of shrinking transistor sizes and the area lim-
itations of photolithography, alternative methods for increasing logic density have
become required to continue meeting the predictions of Moore’s Law [11]. Chiplet
architectures bypass this problem by placing multiple small chips on a single die,
usually in a type of ’2.5D’ configuration [11]. Since these chips need to communicate
with each other, they require an elaborate, dense array of wires to carry information
from chip-to-chip; this is done through the interposer. Each of them also requires a
designated PDN, which will be connected to the PDN for the entire package through
its interposer’s PDN.
2
Figure 1.2: Three-dimensional diagram of a PDN [3].
Figure 1.3: Diagram of a PDN in chiplet architecture [1]. The grey dielectric with
metal layers is the interposer.
3
network. We must also protect the interposer, an asset not present on other ICs.
Moreover, the increased density of busses and power wires means that effects of
localized tampering may be much more damped than in traditional architecture, and
thus much harder to detect.
the methods of transient stimulation and analysis of the resulting signals; we rely upon
both differences in the time domain measurements and differences in the coefficients
of a wavelet transform.
meshes can be grouped into individual sections resembling the interposer and chiplets
in a chiplet system, as shown in figure 4.1. The chiplets will be connected to the
interposer via micro-bump models, while isolated from each other.
automatically parse the output SPICE files and analyze the transient data according
to the methods described in this work.
Our ultimate contribution is the chiplet PDN identification method, with a set
of predictive simulation results for evaluating how our design should be expected to
4
perform if implemented in hardware. We propose methods for that implementation,
as well as possible ways to improve upon the techniques we have developed.
current results of our simulated data, and the methods of analysis we’ve attempted
so far; chapter 5 summarizes the results of hardware experimentation supporting
our work, and their significance for the results we present; chapter 6 provides the
motivation and details of investigations that can continue and improve upon our
methods; chapter 7 concludes this thesis.
5
CHAPTER 2
RELATED WORK
2.1 Modeling
Detailed information on modeling PDN structures is given by Pant and Chiprout
in [3]. This work provided much of the basic information needed for writing the
HSPICE models used for our simulations. Among other contributions, this work
shows that dispersive, 3-dimensional models provide more accurate simulation char-
acteristics than lumped and 2-dimensional models, and that inductances can be ne-
unit cell structures, with the electrical properties parameterized into RLGC values for
SPICE simulation. Likewise, K. Kim et al. offer unit cell modeling methods for on-
chip PDNs and RLGC modeling for through-silicon vias (TSVs) in 3D architectures
6
[5]. This type of modeling is also used by S. Kim et al. and H. Park to create an on-
interposer active decoupling capacitance (decap) network, and optimize placement of
decap, respectively [4, 12]. These works were critical to the structure and parameter
selection of our latest, multi-die models, and therefore to the overwhelming majority
signals, they are capable of identifying friendly wireless nodes from adversarial ones
when combined with Fourier analysis. However, their paper only examined such
techniques over a very limited temperature range (6 ◦ C). Our inspiration to combine
frequency domain transforms with time domain analysis on the original signal came
from Danev and Capkun’s paper.
While the signal processing that must be done for our work may be similar to
that done in [13], there remain significant physical differences between fingerprinting
a wireless sensor node and fingerprinting a chip’s PDN. Mosavirik et al. examine the
impedance and reflectivity of incident transient signals on IC PDNs to characterize
individual chips, classes of chips, and identify location of manufacture [14]. Rather
than the hybrid time domain and transform analysis done by Danev and Capkun,
Mosavirik et al. rely upon a Gaussian Mixture machine learning algorithm to distin-
and measure the voltage response of the PDN, and use the calculated impedance in
the frequency domain to provide the characterization of the PDN. If implemented
7
on FPGA, similar logic would be a good candidate to create and sense the voltage
signals produced in our simulations. While this work is similar to ours in the sense that
with optimal performance and minimum risk of damage. Nonetheless, it is also a good
candidate for security purposes, and is similar to both [14] and the work we have done
with AC analysis (see 4.8.1)
Moini et al. evaluate the performance of ring oscillators (ROs) and large-fanout
flip-flops as power wasters (current sinks), as well as the quality of ROs and time-
to-digital converters (TDCs) for use in detecting power wasting attacks [15]. They
show that ROs are capable of producing current sinks far greater than that of flip-
flops, and that TDCs possess much greater time and voltage resolution than ROs.
More importantly, they show that, when detecting a power wasting attack, TDC
sensors can produce artificially high voltage precision and sample rates of the signal
by shifting the phase of the clock while the power wasting attack is repeated. Should
our fingerprinting be implemented in hardware, this method of TDC sensing may
need to be exploited.
8
CHAPTER 3
There are two models that will be shown in this work. The first is the single die
model, which is the older of the two. This consisted of a single RC mesh for the die,
with an interposer modeled as wires along the edges of the uppermost layer. The
second is the chiplet/multi-die model, which has multiple dies/chiplets (3 in the case
of our tests), each connected to an interposer, but isolated from each other. For this
model, each of the dies and the interposer are modeled as a full mesh RLC grid.
While the tests shown in our work were developed first on the single die model,
the multi-die is a much more accurate model a PDN and much more closely resembles
the chiplet architecture. Therefore, unless otherwise stated, all data presented will be
from the multi-die model. Given that the methods for simulating process variation
(uniqueness) and temperature and supply voltage variation (reliability) are practi-
cally identical for both models, they will be covered solely under the multi-die model
section.
to provide optimal accuracy [3]. For the single die tests, our transient simulations
typically used a 3-layer grid, each with a length and width of 5 nodes (locations where
resistors connect). Each grid has an additional fourth layer as the uppermost layer,
acting as the interposer for the grid. As opposed to the other layers, the interposer
is not a grid, but a square array of wires; it consists of eight sections, each with a
9
resistor and an inductor, and two of these eight sections are connected at each corner
of the next grid layer. In the middle of each interposer edge, a DC voltage source
and a capacitor are added in parallel. The interposer is assumed to have the same
wire dimensions and sheet resistance as the uppermost grid layer, therefore one edge
of the interposer has equivalent resistance to the top grid layer (barring differences
due to the modeled process variation). Likewise, the capacitance for each edge is
equivalent to one full row/column of the grid it’s connected to, up to the process
variation differences.
layers, resembling vias, assume the nominal resistance of the uppermost of the two
layers they are connecting. Like the resistors, each layer also has its own nominal
capacitor values: the top’s is 1.1 f F , the middle’s is 1.0 f F , and the bottom’s is 0.9
f F . Both the nominal resistance and capacitance values are intended to reflect the
shrinking size of the wires as the layers approach the chip’s logic layer. We assume
each inductor in the interposer has inductance of 1.0 nH.
It is important to note that the values used for this model were not calculated,
but merely assumed values that seemed to be reasonable approximations of the target
PDN’s RLC values. This was fixed when we moved to the multi-die model, where we
10
since we are able to change their length and width to produce non-ideal current
sources, similar to what we’d expect to see in a regular chip with logic undergoing
process variation. The drains of the transistors are attached to the desired nodes on
the bottom layer of the grid, the sources are attached to ground, and the sinks are
turned on simply by applying the desired voltage signal to the transistors’ gates.
Our work relies upon two primary signals sent to the NMOS gates: a step pulse-
resembling a workload being turned on and sustained- and a periodic trapezoidal
pulse, resembling a workload being repeatedly turned on, sustained, and turned off.
We assume that the maximum voltage applied to the gate is equal to that supplied
on the interposer.
process variation to components of the PDN system (interposer, single chiplets), and
providing high-variance RLGC components to single nodes in the system. Since we
are also interested in measuring variations on neighboring dice, we must be able to
scripts.
By default, our model is written to provide distinct power and ground planes for
each section of the PDN, though it can be configured to lump the ground RLGC com-
ponents into the power grid to reduce computation time if necessary. Each connection
between nodes follows the unit cell structure, which is a common and computationally
11
Figure 3.1: Diagram of the top-layer grid for the single-die model, with the interposer
connected to its edges.
12
Figure 3.2: Diagram of the bottom-layer grid for the single-die model, with NMOS
sinks. The data here normally assumes these sinks to be in the center.
13
Figure 3.3: Unit cell model of metal layers in a PDN, with transmission lines used
for wires [4]. Note: this particular model does not include a conductive component
(G) in parallel with the capacitor.
Figure 3.4: Diagram of a unit cell, with the ground wire RL values lumped into those
of the the power wires [1]. With explicit power and ground wires, each would have
nominal resistor and inductor values of R and L.
efficient method of modeling PDNs. If there is an explicit ground plane, the node
connections will follow the pattern shown in 3.3. If the ground plane is lumped, the
node connections will follow the pattern given by 3.4. In either case, the unit cells are
stitched together into a grid pattern until the desired number of nodes in the x, y, and
z direction is reached. Our program is capable of providing holes in this grid as well,
where no RLGC connections will be made. This feature can be used to model either
14
For the tests shown in this work, the model used consisted of a chiplet system with
3 dice, connected via interposer. As mentioned in 5, our model lacks a grid for the
package; we would recommend that such a grid would be added to improve accuracy.
Voltage and ground connections are made using through silicon vias (TSVs), modeled
at present as resistors; since the interposer is the uppermost layer, they are connected
to the interposer’s nodes. We assume a TSV pitch of 300 µm. The interposer is
connected to the dice via microbumps, currently modeled as inductors; for these,
we assume a pitch of 150 µm. With minor modifications, the model is capable of
accommodating a more complex RLC model for both the TSVs and microbumps,
but is not configured to do so for our tests. Our model is capable of simulating
multiple ground/power planes with minor adjustments, but we only use one power
and ground plane each for the data in this work.
The RLGC parameters for our unit cells come from [1], which extracted those pa-
rameters from Ansys HFSS simulations of their silicon-based interposer PDNs; those
parameters are summarized in table 3.1. Though [1] doesn’t provide information on
the die-level PDN parameters, we assume that those are the same as the interposer’s
[12]. The unit cell capacitance shown, however, is reduced by a factor of two in our
simulations. This is because it is common for active decap to be used in PDNs [4, 8],
and we expect a reduction in damping to improve uniqueness from the measured re-
sponse to a transient sink. Therefore, we assume that half of the decap in the PDN
could be temporarily disabled until the relevant parts of the identification protocol
are completed. We would, however, recommend that our tests be repeated with the
full decap enabled to ensure that our protocol would be feasible without disabling it.
The TSVs in our model are represented by resistors with a nominal value 40
mΩ, and the micro-bumps by inductors with a nominal value of 1.39 pH [12]. The
15
Figure 3.5: Diagram of the RLGC model for a through-silicon via (TSV) used in [5].
16
TSVs, however, do have significant LC characteristics, resembled by the schematic in
figure 3.5, and we would recommend modifying the simulation to account for these
parameters. We are unaware if there are significant RC characteristics for the micro-
bumps.
Our interposer is modeled with 13 nodes in the x-dimension, and 35 nodes in the
y-dimension. Each of our dice has 11 nodes in the x- and y-dimensions. The space
between nodes in the x-y plane resembles 300 µm spacing on a physical IC, so the
dice are approximately 3 mm2 , and the interposer approximately 10.2 mm x 3.6 mm.
We only use one power and ground plane for the interposer and each of the dice.
Table 3.1: Table of parameters used for our PDN models, taken from [1]. The values
correspond to those in figure 3.4 * nominal capacitance value is reduced by a factor
of 2 to mimic active decap being disabled.
Since our identification methods are reliant upon the process variation intrinsic to
the wires in the PDN, our simulation must be capable of executing Monte Carlo tests.
To do this, we have written our script to generate a unique HSPICE netlist for every
process variation test we wish to run, and have HSPICE run them independently (30
tests/netlist for a single operating corner is typical for our work; the same set of grids
is used at each corner).
Each device in the simulation has its respective values produced from a Gaussian
distribution independently of the other devices in the test using the Python function
17
random.gauss() [17]. However, as we will discuss in section 4.1, we need to be able
to choose which nodes will undergo process variation, and which will remain fixed .
assigned to it for each grid created (30 in our case). Each of these values is assumed
to come from a Gaussian distribution with a standard deviation (1 σ) of 5% relative
to its nominal value. A device that is desired to remain fixed will be drawn randomly
from the same distribution as one that will be varied, but it will be assigned the same
value for each test.
18
CHAPTER 4
The main type of current sink focused on for this work is a single trapezoidal
pulse, resembling an array of power wasters (e.g. ring oscillators) being turned on,
sustained, and then turned off. This type of current sink is easy to produce on and
FPGA, and would require a series of commands that could be executed on the order
[7]- was done in tandem with this project to support its results. Several of the key
parameters used here- namely the aforementioned sink rise, fall and duration times,
and the low sample rate used (1.3 GHz)- come from that work. While we recom-
fingerprinting.
At this point, we should also note that the earliest concept for an identification
protocol tested in our simulations relied upon DC stimulation of a PDN section. It is
our opinion, however, that this provides too intuitive a problem for an adversary to
solve, yields too little uniqueness and would be too sensitive to voltage and tempera-
19
Figure 4.1: Diagram of the dice (sink, suspect, trusted) and interposer (grey) ar-
rangement used for our simulations [6]. Sinks are placed across each point of the sink
die, and measurements are made on the edge near the suspect (blue).
ture variations to be usable for such purposes. Therefore, we will not focus on these
tests in this thesis.
interposer directly connected to it (the die ’shadow’, like that in 1.2). For instance,
this resembles the effects of an attacker de-soldering a desired die from the factory,
and replacing it with a hostile one. The remainder of the PDN has RLGC parameters
taken from a fixed array of random variables, each drawn from the same distribution
as the process variation.
values had greater variation between chips than across environmental (temperature
20
and supply voltage in our models) variations. Thus, our PUF would be able to operate
independently of sensors used to measure environmental variation. However, as shown
in figure 4.8, simulations of our single-die model showed much greater variability over
temperature and supply voltage variations than between chips.
and nominal T&V conditions independently, and determine how much greater that
uniqueness is than the variation caused by thermal noise in the system. Thus, we
have a uniqueness and reliability metric for each corner in our tests. This kind of
enrollment.
The thermal noise in our experiment is modeled as additive white Gaussian noise
(AWGN) with a root-mean square voltage (RMS) given by the equation [20]:
p
vrms = kB T /C (4.1)
intra-class variation, we assume that only the capacitance of the sink/sense die will
contribute to equation 4.1; assuming that the entire interposer and each die con-
tributes to noise reduction, this reduces the amplitude of noise expected from a real
√
chip by a factor of approximately 5. We estimate the resistance contributing to
4.1 by adding the nominal value of each capacitive element on the sink/sense die in
parallel.
21
4.3 Sample Rate
Since any fingerprinting protocol we develop would be limited by the quality of the
sensing hardware on the target system (a problem virtually nonexistent in simulation),
it’s important for us to predict how that hardware will affect the quality of our
ing qualities of detection devices. We will use sample rates of 100 GHz (resembling
an ultra-high quality flash ADC, practically guaranteeing over-sampling of our data),
20 GHz (low-end frequency for a 5G ADC), 1.5 GHz (effective sample rate of a dual
TDC measurement system) and 750 MHz (sample rate of a single TDC in our tandem
hardware experiments).
we run our tests at, we will round our simulated data to 3 different quantization values:
500 µV, 100 µV and 10 µV.
to first evaluate the possibility of using our signal data in its time domain repre-
sentation to identify our devices. We have two methods of identifying a signal: the
first (L1 metric) is to find magnitude of the difference between the measured signal
22
Figure 4.2: Waveforms from 4 different chips in response to a trapezoidal current
sink. Response is shown for a sensor node (blue) as well as a node on the trusted die
(green), directly adjacent to the suspect (see figure 4.1). Current sink varies from 1
mA to 200 mA.
23
Figure 4.3: Zoomed-in view of the response to the current sinks’ rising edge, shown
in figure 4.2.
24
Figure 4.4: Zoomed-in view of the response to the current sinks’ falling edge, shown
in figure 4.2. Notice the variation becomes more noticeable here; this is where we see
most variation in our wavelet transforms.
25
Time Domain L1 Norm Performance 100 GHz 20 GHz 1.3 GHz
10 µV 1.0/0.017 0.999/0.013 0.982/0.022
100 µV 0.998/0.019 0.996/0.02 0.956/0.009*
500 µV 0.996/0.018 0.989/0.021 0.339/0.0*
Table 4.1: True positive rate/false positive rate of the time domain L1 norm identifi-
cation method, given as a function of sample rate and voltage measurement precision.
Rates given as cumulative across temperature and supply voltage corners. 1.0 indi-
cates a 100% rate, and 0.0 a 0% rate. *: indicates points where fingerprinting seems
to fail.
Table 4.2: True positive rate/false positive rate for time domain L2 norm. Same
format and method as 4.1. *: indicates points where fingerprinting seems to fail.
and the factory reference at each point in time the signal is measured, then sum the
differences for all points in time for a single distance metric; the second (L2 metric)
is nearly identical, except that the difference for each time stamp is is squared. The
procedure used for evaluating the L1 and L2 distance metrics for their quality as a
fingerprint is the same as that used for the wavelet transforms, as described in section
4.6.
characteristics in time [21]. Since the variation in the RLC parameters of the PDN
is expected to change the time domain characteristics of our data- such as delay,
rise/fall time, and overshoot/undershoot- wavelet transforms would appear to be good
26
Figure 4.5: PDN response to the rising edge of the sink, plotted at various temperature
and supply voltage conditions. Gathered from the single-die model.
candidates for identifying unique traits of our signal. The wavelet transforms done
coefficients, with each entry corresponding to a particular time and frequency. Our
analyses are made by making comparisons between the CWT matrix of a control
grid and a lineup of other CWT matrices. As mentioned before, each of the chips
in our lineup (including the control grid) is simulated at each of the three tempera-
ture/supply voltage corners. To find uniqueness, we compare the control grid’s signal
to each of the other grids’ at each operating corner. Similarly, to find reliability, we
alter the control grid’s signal with thermal noise from equation 4.1, and compare the
’factory’ measurement (first signal in the lineup) with the other 29 variations of the
friendly signal. Like the uniqueness test, this process is repeated for each tempera-
ture/voltage corner. To ensure that outliers are accounted for and to produce better
convergence of our uniqueness and reliability metrics (true positive/false positive
27
Figure 4.6: Differential wavelet transform between two chips at nominal T&V, across
varying sample rates. Notice the aliasing that occurs as sample rate is decreased.
28
Figure 4.7: Differential wavelet transform between two chips at nominal T&V, across
varying voltage quantization values. Notice the relatively low change in variation.
29
rates), each test (uniqueness and reliability) is repeated for each chip in the 30-chip
sample. That is, each chip will have a friendly signal compared to the remaining 29
frequency axes, then find the difference between the two coefficients at each time-
frequency location. These differences are then summed together to create an L1 sum
of the coefficient differences, which is our uniqueness metric for the wavelet transform.
Based on the variation in L1 sum observed for both the reliability and uniqueness
tests at the given environmental conditions, we set a threshold value for the difference;
an L1 sum greater than the threshold value results in a rejection (hostile), while a
sum less than it results in an acceptance (friendly). The same method is used for the
L2 metric; the only change is that the each difference between coefficients is squared.
Given a spread of time-domain measurements across chips (uniqueness/between-
ation, the optimal outcome for one of these histograms would be a within-class L1
distribution centered near 0 and a between class distribution far from 0, with no
overlap. This is similar to what we see in figure 4.11
reliability at each corner of operation (fast, nominal, slow). Likewise, the quality
of the simulated ADC/TDC (i.e. sample rate and voltage resolution) will change
the relative uniqueness and reliability of the fingerprinting. Therefore, the decision
30
CWT L1 Norm Performance 100 GHz 20 GHz 1.3 GHz
10 µV 0.999/0.013 0.998/0.021 0.96/0.016*
100 µV 0.999/0.02 0.992/0.013 0.871/0.011*
500 µV 0.995/0.017 0.982/0.017 0.339/0.0*
Table 4.3: True positive rate/false positive rate of the CWT L1 norm identification
method, given as a function of sample rate and voltage measurement precision. Rates
given as cumulative across temperature and supply voltage corners. 1.0 indicates a
100% rate, and 0.0 a 0% rate. *: indicates points where fingerprinting seems to fail.
Table 4.4: True positive rate/false positive rate for CWT L2 norm. Same format and
method as 4.3. *: indicates points where fingerprinting seems to fail.
Since we assume that within-class variation is coming from thermal noise in the
PDN, we assume that a manufacturer would be able to run the identification protocol
multiple times on a single board, and produce a distribution for that chip’s within-
class distance metric. Additionally, since we know that the within-class variation
comes from additive white Gaussian noise, we assume that the within-class variation
will be a Gaussian distribution. Finally, we want to make sure that the vast majority
of within-class signals pass the distance threshold, since we want to avoid disposing
of desired devices. Therefore, we set the threshold as the mean of the within-class
distance plus two standard deviations of the same distance; this should ensure that
roughly 98 % of desired chips pass. Any distance above this threshold is assumed
to be associated with a between-class signal, and anything below with a within-class
signal.
31
Figure 4.8: Sum of the L1 differences in wavelet transform coefficients for the unique-
ness (red, inter-chip) tests and reliability (blue, temperature and supply voltage). On
average, the L1 differences are larger for the reliability tests than the uniqueness tests,
indicating that this is a poor metric on its own to identify one of our simulated PDNs.
The blue box at 0 indicates the result for the control grid re-simulated at nominal
temperature and voltage during the reliability tests. Data collected from old, single
chip model.
32
Figure 4.9: Sum of L1 wavelet coefficient differences, 30-chip uniqueness tests.
Figure 4.10: Sum of L1 wavelet coefficient differences for a single chip with 30 different
noise variations added.
Figure 4.11: Uniqueness (top) and reliability (bottom) performance of the wavelet
transform for nominal T&V conditions, with the chips’ signal sampled at 20 GHz
and measured voltage quantized to 100 µV . Note that the minumum L1 value from
process variation is greater than the maximum from the noise variation, so this is a
unique and reliable system at these conditions.
33
Figure 4.12: Results of HSPICE linear sweep of AC frequencies near predicted reso-
nance frequencies of 100 chips. Gathered from the single-die model.
34
4.8 Other Identification Techniques
4.8.1 AC Analysis
in sections 4.5 and 4.6, we evaluated the possibility of using the resonance frequency
of a PDN to identify devices. Though this was only done on our single-die models,
and so have no data on how it would perform in the 3-die chiplet system, it showed
high uniqueness and reliability in our simulations. As shown in figures 4.12 and 4.13,
the vast majority of the chips in a 100-chip spread had a unique resonance frequency
and amplitude. Moreover, as figure 4.14 shows, there was no indication of change in
resonance frequency across temperature and voltage conditions, despite the change
in amplitude.
There are at least two reasons why resonance frequency characterization- or AC
sides of a potentially hostile chiplet, as shown in 4.15: this would open the possibility
of detecting a hostile die by producing a sink on one die, and measuring the response
with a sensor on the other.
vice. Much like our work, this required frequency-swept AC signals to determine the
frequency-domain scattering behavior.
35
Figure 4.13: Scatter plot of the resonance peaks shown in figure 4.12.
36
Figure 4.14: AC frequency sweep of a single chip, with temperature and supply voltage
varying on each iteration. As expected, the resonance frequency remains constant,
while the amplitude changes. The first chip from the tests in figure 4.12 was used for
these tests.
Figure 4.15: Scenario with trusted dice (sink and sensor) on opposite sides of a
potentially hostile die (suspect) [6].
37
4.9 Multi-die measurements
As well as the arrangement used in our 3-die models- a single die with sink and
sensing logic, depicted in figure 4.1- we attempted to use a single die in the system
for sensing and another for generating current sinks, as shown figure 4.15. Instead of
the wavelet transform and time domain methods described in sections 4.5 and 4.6,
we attempted to identify chips by measuring specific time domain features, namely
50-50 propagation delay. In principle, this is nearly identical to the PUF metric used
in [16], but done across PDN planes rather than individual wires. Additionally, we
examined the possibility of introducing a ranking system for delay paths to improve
statistical robustness and reliability across temperature variations. The delay paths
began at a node on the sink die adjacent to the suspect, and ended on the node on
the same y-axis on the sensing die, also directly adjacent to the suspect.
However, our simulations showed fundamental problems with this type of metric.
As shown in figure 4.2, the sensing die experienced a much lower drop in voltage
than the sink die: about 5 µV maximum. This near-negligible drop resulted in very
low uniqueness, and would be extremely difficult for a sensor to measure. Therefore,
38
CHAPTER 5
HARDWARE EXPERIMENTATION
hardware for validation and determining critical parameters. Additionally, the current
iteration of our model has shortcomings not present on an actual board. For example,
as mentioned in section 3.4, our simulations lack a layer for the package, so it is
possible that there are effects on the uniqueness and reliability of our system that
are not observed in our data. It is also possible that the added capacitance and
inductance from the package would lower the frequencies of our PDNs’ harmonics,
which would make our system more effective with lower-sample rate ADCs/TDCs
(see the drop off in performance in tables 4.1, 4.2, 4.3 and 4.4). Hardware, of course,
doesn’t have these shortcomings, and so will give us a more accurate representation
simulated methods on an actual device. This section will be spent discussing Cohen
Hoffing’s findings regarding sink and sensing techniques, and their incorporation into
our models [6].
The board used for evaluating our fingerprinting protocols is the Virtex Ultra-
Scale+ FPGA [7], shown in 5.1. Like our models, this is a 3-die chiplet system.
39
This should ensure that we compare the hardware results to a sufficiently similar
simulation analog.
of current [15], the total current draw for the die will be approximately 20A. These
ROs are turned on sequentially, and it takes about 7 ns for all of them to reach
their maximum current draw, as well as to reach their minimum current draw. These
numbers are assumed for total current draw and rise/fall time in our simulations.
For sensing, a single TDC is placed on the same die as the ROs, at the center of
the edge toward the middle die in the system. The arrangement of the RO sinks and
the TDC is identical to that shown in 4.1. The TDC samples at 600 MHz, which is
approximately the maximum rate for the TDC clock found without producing timing
violations. Experiments have been done with using 2 TDCs with phase-shifted clocks
to measure the PDN response signal and create an artificially higher sample rate, but
have yet to be successful.
Since TDCs produce a thermometer code, with resolution determined by the ham-
ming weight of the measured value from the baseline measurement, we use the max-
imum hamming weight from the hardware tests to determine an approximate lower
bound on the resolution of a TDC. In the experimentation, the maximum hamming
weight observed is about 21 bits. Assuming that the same maximum voltage drop
is observed on the physical board as in our simulated board- about 15 mV- we can
determine the voltage resolution of a TDC to be about 700 µV . This is close to
known resolutions of TDCs [15], as well as the lower bound chosen for our voltage
quantizations. However, we assume a lower bound quantization of 500 µV . In order
40
Figure 5.1: The Virtex UltraScale+ FPGA used for the supporting hardware imple-
mentation [7].
41
CHAPTER 6
Our simulations show high uniqueness and reliability for our wavelet transform
and time domain analyses, and successful operation at all but the worst-case sensing
conditions studied. Given this success, the relatively lightweight nature of our iden-
tification method, and the ubiquity of PDNs in chiplet and monolithic ICs, there is
ample reason to test its performance on hardware and evaluate sensing techniques
for their suitability to our system. However, there remain unexplored methods that
could further improve our system’s uniqueness and reliability, as well as options to
improve our modeling techniques. Therefore, we will document these in this section.
While the parameters we use for our simulation come from literature ([1, 12]), our
in [1], there are several advantages to this: the first is that we would be able to
determine individual resistances for same-layer (x-y) metal connections, and between-
layer (z) metal vias, as the geometries are not necessarily equal; the second is that (if
applicable) we will be able to take thermal expansion and contraction into account
when determining change in resistance, rather than relying upon a fixed percentage
change; the third is that we can model variation in RLGC characteristics with changes
42
Figure 6.1: Lumped model of an IC’s (not chiplet) PDN, with workload [8].
While voltage was supplied to our interposer grid with ideal voltage sources, a real
board would use an active regulator to maintain a steady voltage supply. Additionally,
as mentioned in section 3.4.1, active decap is commonly used in PDNs [4, 8]. In order
to make our simulations more accurate, op-amp models for the regulators and decap
can be added.
evaluate our PDNs in both the time and wavelet domain. Our models are effectively
networks of coupled differential equations, with each node interacting with its neigh-
bors. Adding a package network and/or additional chiplets to the model will create
additional interaction between the large scale components of the whole device. Fur-
43
Figure 6.2: Sketch of the droops from the circuit modeled in 6.1 in the time and
frequency domain [8].
ther, as mentioned in [14], self-heating will produce time- and space-varying changes
to the measured signals by changing the RLGC characteristics of the PDN. Since
ML tends to be well qualified for predicting interaction at varying scales and com-
plex behaviors, it is worth searching for ML algorithms that could provide sufficient
identification systems for our fingerprinting techniques.
For an IC’s PDN, each layer (motherboard, package, die) is expected to have its
own droop characteristics, depending upon the RLC values for that layer, as shown
in figure 6.2 [8]. One can use a Fourier transform to observe the frequency at which
these droop responses occur. Since the interposer adds an additional layer to the PDN,
we’d expect it to add another harmonic to the droop response as well. Therefore, we
can examine the possibility of using Fourier transforms to fingerprint the chips’ and
interposer’s droops on a chiplet, and by extension the device.
44
6.3 New Simulation/Experimentation Types
6.3.1 Bayesian Statistics
Since both time domain and wavelet transform analyses showed high true positive
and low false positive rates for identifying the chips in our sample, we can potentially
use Bayesian statistics to improve the probability of rejecting between-class signals
and accepting within-class signals. As mentioned in 2.2, Denev and Capkun found
that combining time domain and frequency domain analyses was necessary to finger-
print transient signals for wireless nodes [13]. Given the similarity of our work, it
stands to reason that combining the statistics of both will improve the performance
should run our protocol on multiple copies of a single grid, each one with a hole ran-
domly placed within it. Given our current parameters, this would resemble detecting
faults on the order of 600 µm2 , caused by malicious tampering. Alternatively, single
nodes can be altered to have RLGC parameters significantly greater or less than the
nominal, resembling a probe or similar tempering method.
45
CHAPTER 7
CONCLUSION
their distributed PDN system, and described methods of wavelet transforms and time
domain analysis to identify the PDN. Between these, our work shows promising results
for both, with the data typically showing higher uniqueness and reliability for the time
domain. The results we produced give us ample reason to believe that this work ought
to be further evaluated in simulation and experimentation to validate its performance
as a fingerprinting protocol. Developing sensors to operate at the conditions outlined
here will likely be the greatest challenge of implementing our system, though it is
possible that our work overestimates the requirement of its sensors due to the lack
of a package model in our tests. We also suggested additional methods of PDN
fingerprinting and techniques to fine-tune the procedure used for this work.
In the process of this work, we have produced a series of Python scripts capable
of automatically generating and testing the chiplet and interposer PDN structures
for this identification system in HSPICE. We have designed this to provide a thor-
ough, predictive HSPICE model capable of guiding development of such identification
techniques, and changes in efficacy with changes in technology.
Given the success of our system in hardware, we will provide a novel, lightweight
design compatible with virtually any chiplet architecture, and likely monolithic archi-
tectures. Secondarily, hardware experimentation offers the opportunity to verify the
quality of our models. If validated experimentally, our model will have the capability
of serving as a method of automated, rapid testing for chiplet PDN designs.
46
BIBLIOGRAPHY
47
[9] U. Ruhrmair and D. E. Holcomb, “PUFs at a glance,” in Design, Automation
& Test in Europe Conference & Exhibition (DATE), 2014. Dresden,
Germany: IEEE Conference Publications, 2014, pp. 1–6. [Online]. Available:
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6800561
[11] “3 Ways 3D Chip Tech is Upending Computing,” Mar. 2022. [Online]. Available:
https://spectrum.ieee.org/amd-3d-stacking-intel-graphcore
[12] H. Park, J. Park, S. Kim, K. Cho, D. Lho, S. Jeong, S. Park, G. Park, B. Sim,
S. Kim, Y. Kim, and J. Kim, “Deep Reinforcement Learning-Based Optimal
Decoupling Capacitor Design Method for Silicon Interposer-Based 2.5-D/3-D
ICs,” IEEE Transactions on Components, Packaging and Manufacturing
Technology, vol. 10, no. 3, pp. 467–478, Mar. 2020. [Online]. Available:
https://ieeexplore.ieee.org/document/8985416/
[16] A. Deric and D. Holcomb, “Know Time to Die – Integrity Checking for
Zero Trust Chiplet-based Systems Using Between-Die Delay PUFs,” IACR
Transactions on Cryptographic Hardware and Embedded Systems, pp. 391–
412, Jun. 2022. [Online]. Available: https://tches.iacr.org/index.php/TCHES/
article/view/9706
[18] N. H. E. Weste and D. M. Harris, CMOS VLSI design: a circuits and systems
perspective, 4th ed. Boston: Addison Wesley, 2011, oCLC: ocn473447233.
48
[19] R. Groves, D. Harame, and D. Jadus, “Temperature dependence of Q and
inductance in spiral inductors fabricated in a silicon-germanium/BiCMOS
technology,” IEEE Journal of Solid-State Circuits, vol. 32, no. 9, pp. 1455–1459,
Sep. 1997. [Online]. Available: http://ieeexplore.ieee.org/document/628763/
[20] R. Sarpeshkar, T. Delbruck, and C. Mead, “White noise in MOS transistors and
resistors,” IEEE Circuits and Devices Magazine, vol. 9, no. 6, pp. 23–29, Nov.
1993.
[21] M. Frisch and H. Messer, “The use of the wavelet transform in the
detection of an unknown transient signal,” IEEE Transactions on Information
Theory, vol. 38, no. 2, pp. 892–897, Mar. 1992. [Online]. Available:
http://ieeexplore.ieee.org/document/119748/
49