IEEE Circuits and Systems Magazine - Q1 2023 - 37

Abstract
The power density in modern Integrated Circuits (ICs) is tremendous.
For example, Multi-Processor Systems-on-Chip (MPSoCs)
nowadays undergo temperature swings of 40 degrees in 100
milliseconds or less, with rapidly emerging and vanishing submillimeter
hot spots. As such, not only a simulation-based cooling
assessment is vital, but one has to simulate the on-chip thermal
phenomena jointly with the heat dissipation system - historically,
a challenge. In recent years, however, the idea of coupling traditional
3D chip simulators with heat dissipation models written in
Equation-Based Modeling (EBM) languages has proven to be a
game changer. EBM languages allow one to compose a model
by assembling components described in terms of Differential and
Algebraic Equations (DAE) and have the simulation code generated
automatically. In this article, we take a tutorial viewpoint on
the matter just sketched, to put the reader in the position of exploiting
the above technology. We also present the first nucleus
of a model library for cooling systems, that we release as free
software for the scientific and engineering community.
I. Introduction
T
he unprecedented power density of modern ICs,
such as MPSoCs, has created the need for a new
generation of heat dissipation systems. These include
a plethora of technologies, such as liquid and twophase
cooling, Peltier elements, evaporative systems,
and more. Also, the elements just mentioned are often
combined, composing multi-physics cooling solutions.
The relevance of the entailed problems can be appreciated
by looking for example at recent works in the MPSoC
domain, such as [1], [2], [3], and [4].
Since an incorrect behavior of the cooling mechanism
can nowadays have a profound impact on the performance
and reliability of a computing system, simulation-based
heat dissipation assessments are mandatory.
Given the fast dynamics of the involved thermal phenomena
[5], such assessments require to simulate the
said phenomena together with the cooling system, as
well as with the involved thermal/power/performance
policies aboard the chip, see [6] and [7]. Because of this
new scenario, modeling and simulation of heat dissipation
systems need a qualitative leap from several viewpoints,
most notably computation speed, accuracy, and
model maintainability.
The order to fulfill is tall. The problems to address
are inherently cyber-physical, owing to the presence of
hydraulics and thermodynamics [8] (physical) jointly
with cooling circuit control algorithms and on-chip policies
(cyber). Moreover, the physical part is always multidomain,
and shows so wide a variety of configurations
to make a component-based approach mandatory. Also,
the dimension of the systems to simulate can be large,
as fine-grained spatial resolutions may be in order. At
the same time, finally, the level of modeling detail must
be scalable, to achieve the maximum computational efficiency
in each simulation study.
Indeed, traditional thermal simulation approaches
are unsuitable for such new IC operating scenarii. These
approaches are based on exploiting the peculiarities of
dynamic thermal modeling when applied to the case of
ICs. As a result, they do achieve fast simulation, but at
the deliberate expense of modeling generality: the faster
a simulator conceived this way runs, the narrower the
set of cases it can represent is. Modern alternatives
such as Equation-Based Modeling (EBM) suffer from the
symmetric problem: they naturally lend themselves to
representing the heterogeneous physics encountered
when characterizing the transient thermal behavior of
cooling systems but pay for this capability in terms of
computational efficiency.
Recently, we tried to join the best of the two approaches
just mentioned. We made the well-assessed
3D-ICE thermal simulator capable of performing cosimulation
with objectoriented, equation-based modeling
and simulation tools [9]. As such, IC designers can
now build simulators in which 3D-ICE takes care of the
chip thermal model and of interfacing with thermal policies,
while the cooling system model is assembled on
a per-component basis, where components can be described
in an equation-based manner. We coupled the
above modeling approaches in 3D-ICE 3.0 through the
Modelica language [10], [11] and the Functional Mock-up
Interface (FMI) standard [12], [13]. However, the underlying
ideas are general with respect to the used tools.
Coming to the goal of this article, the evidenced revolution
in the IC-thus MPSoC-cooling scenario also
has a particularly important cultural consequence. In
the past, modeling knowledge was practically needed
only on the part of simulation tool developers; for example,
to use 3D-ICE, an MPSoC designer just had to
compile configuration files about the chip design, which
requires no knowledge about heat model equations and
their solution procedure. Now, this is not fully applicable
anymore. Not only has the cooling physics become
very heterogeneous, but also individual cooling systems
are different from one another. This makes it impossible
to provide a " standard " model to tailor with configuration
files and would require the analyst to set up a
Federico Terraneo, Alberto Leva, and William Fornaciari are with the Dipartimento di Elettronica, Informazione e Bioingegneria, Politecnico di Milano,
20133 Milan, Italy (e-mail: federico.terraneo@polimi.it; alberto.leva@polimi.it; william.fornaciari@polimi.it).
David Atienza is with the Embedded Systems Laboratory, École Polythechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland (e-mail: david.
atienza@epfl.ch).
FIRST QUARTER 2023
IEEE CIRCUITS AND SYSTEMS MAGAZINE
37

IEEE Circuits and Systems Magazine - Q1 2023

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q1 2023