You are on page 1of 14

Chapter 14

Thermal Analysis of VLSI System using Successive Over Relaxation (SOR)


Method

Elayaraja Aruchunan, Zailan Siri, Kohilavani Naganthran, Muhamad Hifzhudin Bin Noor
Aziz, Shahirah Akma Ghazali, Mohana Sundaram Muthuvalu, and Jumat Sulaiman

Nowadays, as the size of VLSI is getting smaller, thermal profile is playing an important role
in the system since the high temperature can affect the performance of the system. There is a
various of method in determining the thermal profile of VLSI system. The focus of this study
is to use CN method in discretization of the heat equation to obtain the linear system. The linear
system will be solved by using iterative methods, GS and the proposed SOR method. Then, a
comparison of the efficiency between these two methods will be analysed based on iteration
counts, computer proses time and the maximum temperature. The numerical results show that
SOR is more efficient compared to GS. Therefore, the results for this study may be beneficial
for the future research in solving numerical iterative method.

14.1 Introduction

Very Large-Scale Integrated (VLSI) system is one of the main components in innovation of
the advanced technologies. It is a process of embedding millions of metal oxide-silicon (MOS)
transistors onto a single chip to create an integrated circuit (IC) such as microprocessor,
microchips, etc. As the system is getting advanced, there is some flaw in the system which is
the temperature is one of the main problems that needs to be concerned. The execution, the
speed and the expectancy of the framework from time to time could be reduced due to the high
temperature produced in the system [1]. As a result, in order to retain the system's feature and
capability, it is critical to establish the correct thermal profile of the system, which is one of
the most difficult components for designers to determine as VLSI circuits get more
complicated. The designers were able to obtain the thermal profile of the system at the time by
utilising numerous sensors on the chip and comparing the output to the reference voltage
known as the 'overheating level.' Though, currently, there are numerous methods for obtaining
a thermal profile, and it is dependent on the procedures used to achieve the best results [2].

Furthermore, as the interconnect resistivity increases with temperature, it leads to poor


power-grid insulation-resistance (IR) drops and longer interconnect resistant-capacitor (RC)
retard, thus causing the executions misfortune, complicating timing and noise analysis. At last,
raised temperatures can shorten interconnect and device lifetimes while package reliability can

1
be extremely influenced by a higher temperature slope [3]. The issue of on-chip thermal
analysis is, it requires the solution of the heat equation, which is portrayed in terms of a partial
differential equations (PDE). Usually, PDEs are solved by discretization or fitting scheme,
which changes the problem to solve a set of algebraic equations that are linear [4].

This study is to discretize function of heat equation in equation (1) using Crank-Nicolson
(CN) method and form a linear system. Then, we will apply Successive Over Relaxation (SOR)
method in solving the generated system of linear equation from the heat equation. Lastly, the
objective of this study is to analyse and compare the performances of SOR method with GS
method in solving the generated system of linear equation. Specifically, the remaining study
components are organized as follows: Section 1 contains the brief about the VLSI system, heat
diffusion equation and the objectives described in the detail; Section 2 provides brief
explanations about the CN method to discretize the heat equation, GS method and SOR method
to solve the generated system of linear equation; Section 3 will show the numerical approach
for the heat diffusion equation by discretization using the CN method. Then, the equation
obtained from the discretization is represented in a matrix form to be solved by the GS and
SOR iterative method. Further, the algorithm of the GS and SOR iterative methods are shown;
Section 4 shows some numerical results and discussion based on the results of the numerical
experiments are shown. Solving the thermal issues allows us to analyse the performance of
existing CN-GS and proposed CN-SOR approaches; Section 5 presents the percentage
reduction of the proposed method to validate its efficiency; Section 6 outlined the summary of
the study conclusion.

14.2 Crank-Nicolson method

According to [1], to express the heat conduction in a chip and determine the temperature
profile, we will use the heat diffusion equation as in below
𝜕𝑇(𝑟,
⃗⃗ 𝑡)
𝜌𝐶𝜌 = ∇ ∙ [k(𝑟,
⃗⃗ 𝑇)∇𝑇(𝑟,
⃗⃗ 𝑇) + 𝑔(𝑟,
⃗⃗ 𝑇)] (1)
𝜕𝑡

subject to the Robin’s boundary condition which is the general thermal convection boundary
condition as follows
𝜕𝑇(𝑟,
⃗⃗ 𝑡)
𝑘(𝑟,
⃗⃗ 𝑇) ⃗⃗ 𝑇) = 𝑓𝑖 (𝑟,
+ ℎ𝑖 𝑇(𝑟, ⃗⃗ 𝑇) (2)
𝜕𝑛𝑖

In equation (1), 𝑇, 𝜌, 𝐶𝜌 , 𝑘 and 𝑔 represent temperature, density of material, mass heat capacity,
thermal conductivity and heat energy generation respectively. Meanwhile, in Eq. (2), 𝑛𝑖 , ℎ𝑖 and
𝑓𝑖 (𝑟,
⃗⃗ 𝑇) represent outward direction normal to the boundary condition 𝑖, coefficient of heat
transfer and arbitrary function on the boundary surface respectively [5].

14.3 Numerical Crank-Nicolson method

CN method was used numerically to solve the heat equation by John Crank and Phyllis
Nicolson. It was found to control the instability, as well as to extend the effectiveness and the
accuracy of the implicit and the explicit method [6]. Furthermore, it is necessary to solve a
large linear system. The Bender-Schmidt formula, there is a CN method which will give us an
implicit function

2
−𝜆𝑢𝑖−1,𝑗+1 + 2(1 + 𝜆)𝑢𝑖,𝑗+1 − 𝜆𝑢𝑖+1,𝑗+1 = 𝜆𝑢𝑖−1,𝑗 + 2(1 − 𝜆)𝑢𝑖,𝑗 + 𝜆𝑢𝑖+1,𝑗 (3)

Equation (3) converges on all values of 𝜆, when 𝜆 = 1, that is 𝑘 = 𝑎ℎ2 . The simplest form of
the formula is given by the value of 𝑎 which is average of values of u at 𝑏, 𝑐, 𝑑 and 𝑒 as the
Figure 1 below.

Figure 1: The simplest form of the formula

In this method, the values of 𝑢 at the timestep are obtained by solving a system of linear
equations in the unknowns 𝑢 and 𝑖 as the equation (4) below
1
𝑢𝑖,𝑗+1 = [𝑢 + 𝑢𝑖−1,𝑗+1 + 𝑢𝑖−1,𝑗 + 𝑢𝑖+1,𝑗 ] (4)
4 𝑖+1,𝑗+1

Since the heat diffusion equation in equation (1) is an explicit function, so this method will
be used to discretize the equation to form a linear system. The aim of this method is to put the
all the unknown variables at 𝑛 + 1 on the right-hand side and the known variables at 𝑛 on the
left-hand side. Note that, we need to concern for two dimensional while using this method since
we need to consider for both 𝑖 and 𝑗. In the end, the heat diffusion equation in equation (1) can
be discretized and written in the form of approximation equations. These approximations can
be modelled as a linear system of equations before they can be solved using a linear solver.
Usually, a system of linear equations can be described in matrix notation as equation (5) follows
𝐴𝑥 = 𝑏 (5)

where 𝐴 is a square matrix, 𝑥 and 𝑏 are 𝑛 × 1 vector. Note that, matrix 𝐴 and 𝑏 are given while
𝑥 is an unknown solution which needs to be determined. CN method could be a good choice
for solving the heat equation because it is unconditionally stable for both one-dimensional (1D)
and two-dimensional (2D) applications. This characteristic makes it ideal for any system
involving a conservation law.

3
Figure 2: Description of discretization of 2D heat equation
Figure 2 above is from [7] (Demmel 1996) representing the view of 2D heat equation
on a square plate. Then, to solve the linear system equation obtained from the CN method
above, iterative method is used. Iterative method is a mathematical way of solving a problem
which creates a sequence of approximations. This method is applicable for both linear and
nonlinear problems with huge number of variables. The word iterative or iteration refers to the
technique that solve any linear system problems with successive estimation at each step.
Iterative methods develop series of solution approximation and improve the linear system
solution including large sparse system. Even so, it is not always applicable as there are
convergence criteria that got to be achieved before applying the concepts [1] (Rashid 2016).
As the convergence and the number of iterations depend on the initial guess and the step size,
these two components have been main considerations in any iterative solutions. Iterative
method consists of two methods which are stationary iterative method and non-stationary
iterative method. Stationary iterative method performs the same operations in each iteration on
the current iteration vectors. Meanwhile, non-stationary iterative method is the iterative method
that has iteration-dependent coefficients. GS and SOR method are examples of the stationary
iterative methods that are often being used to solve any linear systems.
14.4 Gauss-Seidel method

GS method is one of the iterative methods to solve linear equations with the unknown variables.
This method is exceptionally straightforward and used in digital computers for computing. GS
is approximately like Jacobi method, except that it uses new values as soon as they are
available. Generally, if the Jacobi method converges, then the GS method will converge faster
than the Jacobi method, though still relatively slowly. Assume the linear system of equations
(5) are examined one at a time in a sequence, and the previously calculated results are used as
soon as they are available, we obtain the GS the equation (6) below
𝑖−1 𝑛
(𝑘+1) 1 (𝑘+1) (𝑘)
𝑥𝑖 = [𝑏 − ∑ 𝑎𝑖𝑗 𝑥𝑗 − ∑ 𝑎𝑖𝑗 𝑥𝑗 ] (6)
𝑎𝑖𝑖 𝑖
𝑗=1 𝑗=1+1

The two important characteristics of the GS method that should be noted are the
computations show up to be serial. Since each component of the new iteration depends upon
all previously computed components, so the updates cannot be done at the same time. Secondly,

4
the new iteration 𝑥 (𝑘+1) depends upon the sequence in which the equations are examined. If
this sequence is changed, the components of the new iterations, not just their order will be
change too. In term of matrices, the definition of GS method can be expressed as

𝑥 (𝑘+1) = (𝐷 − 𝐿)−1 (𝑈𝑥 (𝑘) + 𝑏) (7)

where the matrices 𝐷, −𝐿 and −𝑈 represent the diagonal, strictly lower triangular and strictly
upper triangular parts of 𝐴, respectively. Note that, the GS method is applicable to strictly
diagonally dominant or symmetric positive definite matrices 𝐴.

14.5 Successive Over Relaxation method

After solving the linear equations using GS, we will proceed with the SOR method. According
to [8] (Thevaraja and Light 2016), we use SOR method to solve the set of equations that are
presented in the heat problem. SOR method is derived from the GS method by applying
acceleration parameter 𝜔. It is very similar to the GS method except that it uses the 𝜔 to
decrease the approximation error. For further details, note that for the calculation of 𝑥𝑖 in
equation (6), the variables with index less than 𝑖 are at the (𝑘) iteration while the variables with
index greater than 𝑖 are still at the previous, (𝑘 − 1) iteration [14]. So, the equation for the
SOR method is given as
𝑖−1 𝑛
(𝑘+1) (𝑘) 1 (𝑘+1) (𝑘) (𝑘)
𝑥𝑖 = 𝑥𝑖 + 𝜔 { [𝑏𝑖 − ∑ 𝑎𝑖𝑗 𝑥𝑗 − ∑ 𝑎𝑖𝑗 𝑥𝑗 ] − 𝑥𝑖 } (8)
𝑎𝑖𝑖
𝑗=1 𝑗=1+1

which is it also can be written as the following form


(𝑘+1) (𝑘+1) (𝑘)
𝑥𝑖 = 𝜔𝑥̅ 𝑖 + (1 − 𝜔)𝑥𝑖 (9)

where 𝑥̅ represents a GS iteration and 𝜔 is the extrapolation factor. This extrapolation takes
the form of a weighted average between the previous iteration and the computed GS iterate
successively for each component. In matrix terms, the SOR algorithm can also be written as

𝑥 (𝑘+1) = 𝑥 (𝑘) + 𝜔(𝐷 − 𝐿)−1 (𝑈𝑥 (𝑘) + 𝑏) (10)

here the matrices 𝐷, −𝐿 and −𝑈 represent the diagonal, strictly lower triangular and strictly
upper triangular parts of 𝐴, respectively. The aim is to choose a value for 𝜔 that will accelerate
the rate of convergence of the iterates to the solution. Note that, there will be 5 cases which is
• 𝜔 = 1, the SOR method simplifies to the GS method
• 𝜔 = 0, there is no iteration
• 𝜔 < 1, it is under-relaxation method
• 𝜔 > 1, it is over-relaxation method
• 𝜔 ≥ 2, the equations become divergent

5
14.6 Numerical Approach

CN is used to obtain an algebraic system from the heat equation. The main of this idea is to
obtain a linear system by calculating the average of FDM and BDM. Note that the heat
diffusion equation in equation (1) also can be written as

𝜕𝑇(𝑥, 𝑦, 𝑡) 𝜕 2 𝑇(𝑥, 𝑦, 𝑡) 𝜕 2 𝑇(𝑥, 𝑦, 𝑡)


𝜌𝐶𝜌 = k( + ) + 𝑔(𝑥, 𝑦, 𝑡) (11)
𝜕𝑡 𝜕𝑥 2 𝜕𝑦 2

By using FDM and BDM, we discretize the equation (11) above. The following equation
represent the formula of FDM and BDM of the heat diffusion equation, respectively
𝑛+1 𝑛 𝑛+1 𝑛+1 𝑛+1 𝑛+1 𝑛+1 𝑛+1
𝑇𝑖,𝑗 − 𝑇𝑖,𝑗 𝑇𝑖+1,𝑗 − 2𝑇𝑖,𝑗 + 𝑇𝑖−1,𝑗 𝑇𝑖,𝑗+1 − 2𝑇𝑖,𝑗 + 𝑇𝑖,𝑗−1
= + (12)
∆𝑡 (∆𝑥)2 (∆𝑦)2

and
𝑛+1 𝑛 𝑛 𝑛 𝑛 𝑛 𝑛 𝑛
𝑇𝑖,𝑗 − 𝑇𝑖,𝑗 𝑇𝑖+1,𝑗 − 2𝑇𝑖,𝑗 + 𝑇𝑖−1,𝑗 𝑇𝑖,𝑗+1 − 2𝑇𝑖,𝑗 + 𝑇𝑖,𝑗−1
= + (13)
∆𝑡 (∆𝑥)2 (∆𝑦)2

Then, to proceed with the CN method, we will substitute the average of the FDM and BDM in
equation (12) and equation (13) respectively into equation (11)
𝑛+1 𝑛
𝑇𝑖,𝑗 − 𝑇𝑖,𝑗
𝜌𝐶𝜌
∆𝑡
𝑛+1 𝑛+1 𝑛+1 𝑛+1 𝑛+1 𝑛+1
k 𝑇𝑖+1,𝑗 − 2𝑇𝑖,𝑗 + 𝑇𝑖−1,𝑗 𝑇𝑖,𝑗+1 − 2𝑇𝑖,𝑗 + 𝑇𝑖,𝑗−1
= [ + (14)
2 (∆𝑥)2 (∆𝑦)2
𝑛 𝑛 𝑛 𝑛 𝑛 𝑛
𝑇𝑖+1,𝑗 − 2𝑇𝑖,𝑗 + 𝑇𝑖−1,𝑗 𝑇𝑖,𝑗+1 − 2𝑇𝑖,𝑗 + 𝑇𝑖,𝑗−1 𝑛
+ + ] + 𝑔𝑖,𝑗
(∆𝑥)2 (∆𝑦)2

Let consider ∆𝑥 = ∆𝑦 = ℎ, then equation (14) can be written as


𝑛+1 𝑛
𝑇𝑖,𝑗 − 𝑇𝑖,𝑗
𝜌𝐶𝜌
∆𝑡
k (15)
= (𝑇 𝑛+1 + 𝑇𝑖−1,𝑗
𝑛+1 𝑛+1
+ 𝑇𝑖,𝑗+1 𝑛+1
+ 𝑇𝑖,𝑗−1 𝑛+1
− 4𝑇𝑖,𝑗 𝑛
+ 𝑇𝑖+1,𝑗 𝑛
+ 𝑇𝑖−1,𝑗
2ℎ2 𝑖+1,𝑗
𝑛 𝑛 𝑛 𝑛
+ 𝑇𝑖,𝑗+1 + 𝑇𝑖,𝑗−1 − 4𝑇𝑖,𝑗 ) + 𝑔𝑖,𝑗

Then, simplify the equation (15), and it can be rewrite as

𝑛+1 𝑛 k∆𝑡
𝑇𝑖,𝑗 − 𝑇𝑖,𝑗 = (𝑇 𝑛+1 + 𝑇𝑖−1,𝑗
𝑛+1 𝑛+1
+ 𝑇𝑖,𝑗+1 𝑛+1
+ 𝑇𝑖,𝑗−1 𝑛+1
− 4𝑇𝑖,𝑗 𝑛
+ 𝑇𝑖+1,𝑗 𝑛
+ 𝑇𝑖−1,𝑗
𝜌𝐶𝜌 2ℎ2 𝑖+1,𝑗
𝑛
𝑔𝑖,𝑗 ∆𝑡 (16)
𝑛 𝑛 𝑛
+ 𝑇𝑖,𝑗+1 + 𝑇𝑖,𝑗−1 − 4𝑇𝑖,𝑗 ) +
𝜌𝐶𝜌

Now, we move all the 𝑛 + 1 to the LHS and all 𝑛 to the RHS

6
𝑛+1 𝑛+1 𝑛+1 𝑛+1 𝑛+1 𝑛+1
𝑇𝑖,𝑗 − 𝜆(𝑇𝑖+1,𝑗 + 𝑇𝑖−1,𝑗 + 𝑇𝑖,𝑗+1 + 𝑇𝑖,𝑗−1 − 4𝑇𝑖,𝑗 )
𝑛
𝑛 𝑛 𝑛 𝑛 𝑛 𝑛
𝑔𝑖,𝑗 ∆𝑡 (17)
= 𝑇𝑖,𝑗 + 𝜆(+𝑇𝑖+1,𝑗 + 𝑇𝑖−1,𝑗 + 𝑇𝑖,𝑗+1 + 𝑇𝑖,𝑗−1 − 4𝑇𝑖,𝑗 ) +
𝜌𝐶𝜌

where
𝑘∆𝑡
𝜆=
𝜌𝐶𝜌 2ℎ2
𝑔𝑛 ∆𝑡
𝑖,𝑗
Since 𝜌𝐶 = 0, thus we can ignore it and rearrange equation (17) and rewrite it as
𝜌

𝑛+1 𝑛+1 𝑛+1 𝑛+1 𝑛+1


(1 + 4𝜆)𝑇𝑖,𝑗 − 𝜆(𝑇𝑖+1,𝑗 + 𝑇𝑖−1,𝑗 + 𝑇𝑖,𝑗+1 + 𝑇𝑖,𝑗−1 )
𝑛 𝑛 𝑛 𝑛 𝑛 (18)
= (1 − 4𝜆)𝑇𝑖,𝑗 + 𝜆(+𝑇𝑖+1,𝑗 + 𝑇𝑖−1,𝑗 + 𝑇𝑖,𝑗+1 + 𝑇𝑖,𝑗−1 )

14.6.1 Formula of approximation


In this section, we will approximate the heat diffusion equation using the discretization
schemes as stated in equation 8.
Starts at
𝑖 = 𝑗 = 1,
𝑛+1
(1 + 4𝜆)𝑇11 𝑛+1 𝑛+1 𝑛+1 𝑛+1 )
− 𝜆(𝑇21 + 𝑇01 + 𝑇12 + 𝑇10
𝑛 𝑛 𝑛 𝑛 𝑛)
= (1 − 4𝜆)𝑇11 + 𝜆(𝑇21 + 𝑇01 + 𝑇12 + 𝑇10
𝑖=𝑗=2
𝑛+1
(1 + 4𝜆)𝑇22 𝑛+1 𝑛+1 𝑛+1 𝑛+1 )
− 𝜆(𝑇32 + 𝑇12 + 𝑇23 + 𝑇21
𝑛 𝑛 𝑛 𝑛 𝑛)
= (1 − 4𝜆)𝑇22 + 𝜆(𝑇32 + 𝑇12 + 𝑇23 + 𝑇21
𝑖=𝑗=3
𝑛+1
(1 + 4𝜆)𝑇33 𝑛+1 𝑛+1 𝑛+1 𝑛+1 )
− 𝜆(𝑇43 + 𝑇23 + 𝑇34 + 𝑇32
𝑛 𝑛 𝑛 𝑛 𝑛)
= (1 − 4𝜆)𝑇33 + 𝜆(𝑇43 + 𝑇23 + 𝑇34 + 𝑇32
𝑖=𝑗=4
𝑛+1
(1 + 4𝜆)𝑇44 𝑛+1 𝑛+1 𝑛+1 𝑛+1 )
− 𝜆(𝑇54 + 𝑇34 + 𝑇45 + 𝑇43
𝑛 𝑛 𝑛 𝑛 𝑛)
= (1 − 4𝜆)𝑇44 + 𝜆(𝑇54 + 𝑇34 + 𝑇45 + 𝑇43

𝑖 =𝑗 =𝑁−1
𝑛+1 𝑛+1 𝑛+1 𝑛+1 𝑛+1
(1 + 4𝜆)𝑇𝑁−1,𝑁−1 − 𝜆(𝑇𝑁,𝑁−1 + 𝑇𝑁−2,𝑁−1 + 𝑇𝑁−1,𝑁 + 𝑇𝑁−1,𝑁−21 )
𝑛 𝑛 𝑛 𝑛 𝑛
= (1 − 4𝜆)𝑇𝑁−1,𝑁−1 + 𝜆(𝑇𝑁,𝑁−1 + 𝑇𝑁−2,𝑁−1 + 𝑇𝑁−1,𝑁 + 𝑇𝑁−1,𝑁−2 )
Generally, as we get a large system of linear equations above, so equation (18) can be
represented in matrix form as

𝐴𝑇 = 𝑏 (19)

where
• Matrix 𝐴 is a positive definite, strictly diagonally dominant and tridiagonal matrix
• Matrix 𝑇 is a single column matrix for the unknown values of 𝑇𝑖,𝑗 𝑛+1

• Matrix 𝑏 is a single column matrix for the known values of RHS


The matrix representation is as follows,

7
1 + 4𝜆 −𝜆 0 … 0
−𝜆 1 + 4𝜆 −𝜆 ⋱ ⋮
𝐴= 0 −𝜆 1 + 4𝜆 ⋱ 0
⋮ ⋱ ⋱ ⋱ −𝜆
( 0 … 0 −𝜆 1 + 4𝜆)
𝑛+1
𝑇11
𝑛+1
𝑇22
𝑇= 𝑛+1
𝑇33

𝑛+1
(𝑇𝑖=𝑗=𝑁−1 )
𝑛
(1 − 4𝜆)𝑇11 𝑛 𝑛 𝑛 𝑛)
+ 𝜆(+𝑇21 + 𝑇01 + 𝑇12 + 𝑇10
𝑛
(1 − 4𝜆)𝑇22 𝑛 𝑛 𝑛 𝑛)
+ 𝜆(𝑇32 + 𝑇12 + 𝑇23 + 𝑇21
𝑛
(1 − 4𝜆)𝑇33 𝑛 𝑛 𝑛 𝑛)
𝑏= + 𝜆(𝑇43 + 𝑇23 + 𝑇34 + 𝑇32

(1 𝑛 𝑛 𝑛 𝑛 𝑛
( − 4𝜆)𝑇𝑁−1,𝑁−1 + 𝜆(𝑇𝑁,𝑁−1 + 𝑇𝑁−2,𝑁−1 + 𝑇𝑁−1,𝑁 + 𝑇𝑁−1,𝑁−2 ))

Matrix A is decomposed into 𝐷, −𝐿 and −𝑈 which represent the diagonal, strictly lower
triangular and upper triangular parts of 𝐴 respectively as if

𝐴=𝐷−𝐿−𝑈 (20)

For more detail, this matrix representation of 𝐷, 𝐿 and 𝑈 is shown as follows,


as follows,
1 + 4𝜆 0 0 … 0
0 1 + 4𝜆 0 ⋱ 0
𝐷= ⋮ 0 1 + 4𝜆 0 ⋮
0 ⋱ 0 ⋱ 0
( 0 0 ⋯ 0 1 + 4𝜆)
0 0 0 … 0
−𝜆 0 0 ⋱ 0
𝐿= ⋮ −𝜆 0 0 ⋮
0 ⋱ −𝜆 ⋱ 0
(0 0 ⋯ −𝜆 0)
0 −𝜆 0 … 0
0 0 −𝜆 ⋱ 0
𝑈= ⋮ 0 0 −𝜆 ⋮
0 ⋱ 0 ⋱ −𝜆
(0 0 ⋯ 0 0)
Thus, by substituting equation (20) into equation (19), by GS it is defined by the iteration

(𝐷 − 𝐿)𝑇 𝑛+1 = 𝑈𝑇 𝑛 + 𝑏 (21)

Hence, we can simplify it as

8
𝑇 𝑛+1 = (𝐷 − 𝐿)−1 (𝑈𝑇 𝑛 + 𝑏) (22)

Then, to compare the efficiency of the GS and SOR method, we apply the proposed method
which is SOR to the equation. To proceed with SOR method, we just apply the method by
adding acceleration parameter, 𝜔 to the GS method. Therefore, equation (20) can be expanded
to form SOR general equation as

𝑇 𝑛+1 = 𝑇 𝑛 + 𝜔(𝐷 − 𝐿)−1 (𝑈𝑇 𝑛 + 𝑏) (23)

where 𝜔 is the acceleration parameter, 0 < 𝜔 < 2. Note that when 𝜔 = 1, the equation (23)
will represents the GS method.
14.6.2 Formulation and implementation of Gauss-Seidel method
In general, the algorithm for the existing GS iterative method to solve the linear equation in
equation (22) is given as below

Algorithm 1: GS Iterative method


Step 1 Set 𝑘 = 1
Step 2 While (𝑘 ≤ 𝑁), do Steps 3-6
Step 3 For 𝑖 = 1,2,3, … , 𝑛
(𝑘+1) 1 (𝑘+1) (𝑘)
set 𝑇𝑖 = 𝑎 [𝑏𝑖 − ∑𝑖−1
𝑗=1 𝑎𝑖𝑗 𝑇𝑗 − ∑𝑛𝑗=1+1 𝑎𝑖𝑗 𝑇𝑗 ]
𝑖𝑖

Step 4 If ∥ 𝑇 (𝑘+1) − 𝑇 (𝑘) ∥< 𝜀, then OUTPUT: 𝑇 (𝑘+1)


(STOP)
Step 5 Set 𝑘 = 𝑘 + 1

Step 6 For 𝑖 = 1,2,3, … , 𝑛


(𝑘) (𝑘+1)
set 𝑇𝑖 = 𝑇𝑖
Step 7 OUTPUT: Maximum number of iterations exceeded
(STOP)
where,
𝑛 = number of equations and unknowns
𝑎𝑖𝑗 = entries of Matrix 𝐴 where 1 ≤ 𝑖, 𝑗 ≤ 𝑛
𝑏𝑖 = entries of Matrix 𝑏 where 1 ≤ 𝑖 ≤ 𝑛
𝑇 (𝑘) = 𝑘 𝑡ℎ iteration of 𝑇
𝑇 (𝑘+1) = the next or 𝑘 + 1 iteration of 𝑇
𝜀 = 10−10, error of tolerance
𝑁 = maximum number of iterations

9
14.6.3 Formula and implementation of Successive Over Relaxation method

In this section, the algorithm for the existing SOR iterative method to solve the linear equation
in equation (23) is given as below

Algorithm 2: SOR Iterative method


Step 1 Set 𝑘 = 1
Step 2 While (𝑘 ≤ 𝑁), do Steps 3-6
Step 3 For 𝑖 = 1,2,3, … , 𝑛
(𝑘+1) (𝑘) 1 (𝑘+1) (𝑘)
set 𝑇𝑖 = (1 − 𝜔)𝑇𝑖 + 𝑎 [𝜔(𝑏𝑖 − ∑𝑖−1
𝑗=1 𝑎𝑖𝑗 𝑇𝑗 − ∑𝑛𝑗=1+1 𝑎𝑖𝑗 𝑇𝑗 )]
𝑖𝑖

Step 4 If ∥ 𝑇 (𝑘+1) − 𝑇 (𝑘) ∥< 𝜀, then OUTPUT: 𝑇 (𝑘+1)


STOP
Step 5 Set 𝑘 = 𝑘 + 1

Step 6 For 𝑖 = 1,2,3, … , 𝑛


(𝑘) (𝑘+1)
set 𝑇𝑖 = 𝑇𝑖
Step 7 OUTPUT: Maximum number of iterations exceeded
(STOP)
where,
𝑛 = number of equations and unknowns
𝑎𝑖𝑗 = entries of Matrix 𝐴 where 1 ≤ 𝑖, 𝑗 ≤ 𝑛
𝑏𝑖 = entries of Matrix 𝑏 where 1 ≤ 𝑖 ≤ 𝑛
𝑇 (𝑘) = 𝑘 𝑡ℎ iteration of 𝑇
𝑇 (𝑘+1) = the next or 𝑘 + 1 iteration of 𝑇
𝜔 = weighted parameter, given value
𝜀 = 10−10, error of tolerance
𝑁 = maximum number of iterations

14.7 Numerical Experiments

The heat diffusion equation is discretized by CN method and solved by the GS and SOR
iterative methods. Then the results from these two iterative methods are being compared to
analyze the performances. In this section, the problem is tested using 2 combinations which are
CN-GS and CN-SOR iterative methods. For numerical analysis, there parameters are
considered; number of iterations, computational time in seconds and maximum temperature
(𝑇𝑚𝑎𝑥 ) in Kelvin (𝐾). For convergence test, the tolerance error is set to 𝜀 = 10−10 and the
tested mesh sizes are 16, 32, 64, 128 and 256. The values of the parameters are based on the
property of Silicon which is basically the material used in most VLSI systems. The 2D model
length, 𝐿 is set to 4.75m. The material used in this project is Silicon (Si) since Si can conduct
electricity when the heat is applied. The property of Si is shown as in the Table 1. Note that,
under adiabatic condition, the boundary condition is assigned to zero and initial temperature,
𝑇𝑖𝑛 of the system is set under room temperature as 𝑇𝑖𝑛 = 293.15 𝐾. All the simulations are
performed on computer model type Intel® Core™ i5-7200U 2.5GHz and codes are written in
Borland C++ programming language.

10
Comparison of number of iterations between GS
and SOR iterative methods
60000

Number of Iterations
40000

20000

0
16x16 32x32 64x64 128x128 256x256
Mesh Size

GS SOR

Figure 3: Graph comparison of number of iterations between GS and SOR method

Comparison of computational time in seconds


Computational time (seconds)

between GS and SOR iterative methods


6000

4000

2000

0
16x16 32x32 64x64 128x128 256x256
Mesh Size

GS SOR

Figure 4: Graph comparison of computational time between GS and SOR

T_max vs mesh size for both GS and SOR iterative


method
295.718
295.716
295.714
295.712
295.71
295.708
295.706
295.704
16x16 32x32 64x64 128x128 256x256
Mesh Size
Tmax

GS SOR

Figure 5: 𝑇𝑚𝑎𝑥 vs mesh size for both GS and SOR


Specifically, Table 2 shows the numerical results based on number of iterations,
computational time and maximum absolute errors that are computed by applying GS and SOR
iterative methods with the corresponding CN method discretization to solve the problem.
According to Table 2, the number of iteration and the computational time by SOR iterative
method dropped drastically compared to the GS iterative method. This is due to the weighted

11
parameter which is chosen based on the least number of iterations for each mesh size and
discretization method.
From Table 2, we can plot a line graph as in Figure 3 to show the comparison of number
of iterations between GS and SOR iterative method. As we can see from the graph above, for
every mesh size, SOR iterative method has a least number of iterations compares to the GS
iterative method. Also, the bigger the mesh size, the more the number of iterations. Then, from
Table 2, we can plot a line graph as in Figure 4 to represent the comparison of computation
time in seconds between GS and SOR iterative method. And as we can see above, SOR iterative
method requires less time compared to GS iterative method. Also, the computational time
increases as the mesh size increases. Note that, the least the number of iterations and the smaller
the computational time, the better the performances are. So, both Figure 3 and Figure 4 show
that SOR method brings a better performance in terms of number of iterations and
computational time compared to the GS method. Figure 5 above shows the graph between the
mesh size and the temperature for both GS and SOR methods. This graph is plotted based on
Table 2. From Figure 5, we can conclude that both iterative methods compute the same 𝑇𝑚𝑎𝑥 .
The 𝑇𝑚𝑎𝑥 is increasing as the mesh size increases and starts to drop a little at mesh size
256 × 256. As we can see from Figure 3 and Figure 4 above, SOR method has a least value of
computational results either from number of iterations or computational time compared to the
GS method. Alongside, the output for 𝑇𝑚𝑎𝑥 for both iterative methods are the same.
14.8 Percentage reduction calculation

For this study, the percentage reductions for the number of iterations and computational
time will be examined. To study the performance of the proposed method, a percentage decline
in the number of iterations and computational time of CN-GS and CN-SOR is determined in
the percentage. The percentage reduction for GS and SOR for 16, 32, 64, 128 and 256 mesh
sizes will be tabulated in the next chapter. The following shows how the percentage of
reduction will be calculated,

𝛼−𝛾
ℛ=( ) × 100% (24)
𝛼

where,
ℛ = Percentage of reduction
𝛼 = Number of iterations or computational time for tested GS method
𝛾 = Number of iterations or computational time for tested SOR method
From the computational results in Table 2, we can analyse the percentage reduction
analysis. Percentage of reduction is calculated to show the efficiency of SOR method compared
to GS method. Table 3 shows the results of percentage reduction analysis that have been carried
out from GS to SOR based on number of iterations and computational time. Note that, the
greater the percentage reduction analysis are, the more efficient the method is. Therefore, as
shown from all the figures above, the computational results show that SOR iterative method is
more convenient compared to the GS iterative method
14.9 Conclusion

The outcome and contribution on this study are, we manage to apply the CN method for
the discretization of heat diffusion equation effectively. Based on the research, observed that
the implementation of the discretization method manages to obtain a large system of linear

12
equations. Then, by represent the system of linear equations in matrix form, we successfully
implemented the proposed SOR method and solve the problem. Lastly, we manage to show the
comparison between SOR and GS, which brings that SOR method is better than GS method.
The utmost goal of this study is to apply and explore the performance of SOR iterative method
for the generated large linear system. The GS method was used as a control method to validate
the efficiency of the proposed method. To show the efficiency of SOR iterative method, two
important criteria (number of iteration and computational time) are considered. We also obtain
the maximum temperature which might be thermal profile for the VLSI system for each mesh
size. In conclusion, we manage to fulfill the objectives of this study.
References

[1] Rashid, N. (2016). Thermal Analysis of VLSI System: A Simulation Study. Electrical &
Electronics, Universiti Teknologi PETRONAS. Bachelor of Engineering: 33.
[2] Pedram, M. and S. Nazarian (2006). "Thermal Modeling, Analysis and Management in
VLSI Circuits: Principles and Methods." Proceedings of the IEEE 94(8): 1487-1501.
[3] Huang, W., et al. (2006). "HotSpot: A Compact Thermal Modelling Methodology for Early-
Stage VLSI Design." IEEE Transactions on Very Large Scale Integration (VLSI) Systems
14(5): 501-513.
[4] Heriz, V. M., et al. (2007). Method of images for the fast calculation of temperature
distributions in packaged VLSI chips. Thermal Investigation of ICs and Systems. Budapest,
Hungary.
[5] Wang, T.-Y. and C. C.-P. Chen (2003). Thermal-ADI—A Linear-Time Chip-Level
Dynamic Thermal-Simulation Algorithm Based on Alternating-Direction-Implicit (ADI)
Method. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, IEEE. 11.
[6] Umeorah, N. and P. Mashele (2019). "A Crank-Nicolson finite difference approach on the
numerical estimation of rebate barrier option prices." Cogent Economics & Finance 7(1).
[7] Demmel, J. (1996). Sources of Parallelism and Locality in Simulation (continued).
University of California, Berkeley.
[8] Thevaraja, M. and E. Light (2016). "Applying the Successive Over-relaxation Method to a
Real World Problems." American Journal of Applied Mathematics and Statistics 4(4): 113-
117.

13
Appendixes
Table 1: Properties of Silicon (Si)

Material Properties (Silicon)

Thermal Conductivity, 𝑘(𝑤/𝑚𝐾) 149

Thermal Density, 𝜌(𝑘𝑔⁄𝑚3 ) 2330

Specific Heat Capacity, 𝐶𝜌 (𝑗⁄𝑘𝑔𝐾) 703

Table 2: Numerical results based on number of iterations, CPU time and 𝑇𝑚𝑎𝑥

CPU time
N Methods Number of Iterations 𝑇𝑚𝑎𝑥 (𝐾)
(seconds)
GS 353 0.32 295.708
16 × 16
SOR 45 (𝜔 = 1.6) 0.16 295.708
GS 1266 2.08 295.7138
32 × 32
SOR 114 (𝜔 = 1.8) 0.67 295.7138
GS 4484 28.02 295.7153
64 × 64
SOR 186 (𝜔 = 1.9) 3.85 295.7153
GS 15624 308.32 295.7155
128 × 128
SOR 798 (𝜔 = 1.9) 58.06 295.7155
GS 53146 4922.04 295.7148
256 × 256
SOR 2792 (𝜔 = 1.9) 717.52 295.7148
Table 3: Percentage Reduction Analysis from GS to SOR

Percentage of Reduction
Parameters Mesh Size
16 × 16 32 × 32 64 × 64 128 × 128 256 × 256
Number of
87.2521 90.9953 95.8519 94.8925 94.7465
iterations
Computational
50.0000 67.7885 86.2598 81.1689 85.4223
time(seconds)

14

You might also like