Você está na página 1de 13

OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .

Analysis of the Static Voltage Drop in the Power Distribution


Network of a System-on-chip
Naval Andrianjanandrasana, Martin Gould,
Jean-Charles Seguis and Pawel Szerling,
Contents
1 Introduction 3
2 Statement of The Problem 3
3 The Continuous Approximation 4
4 Solutions for Uniform Voltage Requirements 5
5 Single Block Current Density 9
6 Multiple Block Current Density 10
7 Optimal Chip Congurations 11
8 Conclusions 11
9 Acknowledgements 12
References 12
1
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
2
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
1 Introduction
Day by day, technology is becoming an increasingly central part of life for us all. Whether controlling cinemas,
computers, calculators or cars, the explosive growth of microchip technology during the last fty years has
undoubtedly changed the way that we live. But the technological advances which have allowed ever more
complex circuits to reside in ever smaller spaces have presented manufacturers with an interesting problem:
how to provide power to the chips in a manner which is both ecient and reliable.
The architecture of a microchip is both complex and intricate, and the exact structure of any given chip
depends on the purpose for which it was manufactured. Despite this, the manner in which dierent microchips
are powered is generally very similar: the transistors (which form the basis of functionality in the chip) are
sandwiched between two parallel meshes of thin wires, and a voltage is applied to the boundary of the
upper mesh by some external power source. In order to ease exposition, we name the intersection points in
the meshes nodes. Current ows through the upper mesh, and also ows down to the layer of transistors
via connections from the nodes in the upper mesh. After supplying the transistors with the necessary energy,
the current then ows down to the nodes in the lower mesh before nally owing back to the power source.
A schematic for such a setup is displayed in Figure 1. In the schematic, only one row of vertical connections
is shown in reality, such vertical connections exist at every node.
Figure 1: Schematic of the Supply of Power to Transistors in a Microchip
2 Statement of The Problem
In order for the chip to maintain functionality, guaranteed levels of voltage must be provided to each transistor
that is, each node (x, y) has a minimum voltage requirement V
req
(x, y). Each node (x, y) actually
receives a voltage V
act
(x, y). The basic fesibility constraint for proper operation of the microchip is then:
V
act
(x, y) V
req
(x, y) (x, y) (2.1)
The diculty in the problem comes from the fact that there is a small but unavoidable amount of resistance
present in the mesh. Nodes are separated by a distance of x in the x direction and y in the y direction,
and the wires which run north to south in the mesh have a thickness of dx, while the wires which run east
to west in the mesh have a thickness of dy.
1
In a real microchip, wires in the mesh are packed tightly so
that parallel wires lie very close to each other. The physical constraint on how close these wires can be is,
therefore, the thickness of the wires themselves, and so dx = x and dy = y for our purposes. This setup
is depicted in Figure 2.
1
At this stage, we are not taking innitessimal limits, but merely identifying the relevant dimensions in the physical setup.
3
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
(x, y) (x +x, y)
(x, y+y)
(x +x, y+y)
x
y
dx
dy
Figure 2: Diagram Describing Physical Setup of the Problem
For a variety of reasons (including performance, size and heat-management) manufacturers desire to
minimise the dierence in voltage between the edge of the mesh and any particular node. The manufacturers
therefore wish to solve the problem:
min max
x,y
V
act
(x, y) (2.2)
subject to V
act
(x, y) V
req
(x, y) (2.3)
To avoid endless repetition, we henceforth write V (x, y) for V
act
(x, y).
The structure of a microchip dictates that V (x, y) = 0 for any node (x, y) on the boundary , but deriving
a closed-form expression for V (x, y) is much harder for nodes (x, y) \ . Under the setup depicted in
Figures 1 and 2, we derive a system of equations relating the voltage at any given node in the mesh to the
voltage of its four immediate neighbours. In particular, by Kirchos Law [3]:
V (x, y) =
V (x+x,y)
R(x+
1
2
x,y)
+
V (xx,y)
R(x
1
2
x,y)
+
V (x,y+y)
R(x,y+
1
2
y)
+
V (x,yy)
R(x,y
1
2
y)
+ I(x, y)
1
R(x+
1
2
x,y)
+
1
R(x
1
2
x,y)
+
1
R(x,y+
1
2
y)
+
1
R(x,y
1
2
y)
(2.4)
where I(x, y) and R(x, y) denote the ingoing current and resistance at (x, y) respectively.
3 The Continuous Approximation
Solving the problem described in Equations (2.2) and (2.3) for the function V (x, y) given by Equation (2.4)
is a very diult task in a discrete setting with a specic number of nodes, not least because the addition of
a single extra wire to the mesh creates a large number of new nodes. But as the number of wires grows, and
because the chip size is very small, the system behaves more and more like a plane along which current is
free to ow in both the x and y directions. We now demonstrate how, in this limit that is, the limit as x
and y tend to zero the system can be viewed as being governed not by an enormous system of equations,
but by a single partial dierential equation:

2
V
x
2
1
R
x
+

2
V
y
2
1
R
y
= J
V (x, y) = 0, (x, y)
(3.1)
4
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
Deriving Equation (3.1)
The key concept in the derivation is that of sheet resistance [4]. Given the sheet resistance R
s
of the material
from which the wires are manufactured, the resistance experienced by the ow between a neighbouring pair
of nodes in the x direction (denoted R
x
) and in the the y direction (denoted R
y
) is:
R
x
= R
s
(x, y)
x
y
R
y
= R
s
(x, y)
y
x
(3.2)
Using the expressions from Equation (3.2) in Equation (2.4), and taking the limits x, y 0 yields:
V (x, y) =
(V (x + x, y) + V (x x, y)) R
y
y
x
+ (V (x, y y) + V (x, y + y)) R
x
y
x
+ IR
x
R
y
2
_
R
x
x
y
+ R
y
x
y
_ (3.3)
But notice that the Taylor expansion of V (x x, y) is given by:
V (x x, y) = V (x, y)
V
x
x +
1
2

2
V
x
2
x
2

1
6

3
v
x
3
x
3
+O(x
4
)
So:
V (x + x, y) + V (x x, y) = 2V (x, y) +

2
V
x
2
x
2
+O(x
4
) (3.4)
Then, substituting Equation (3.4) (and a similar expression for y) into Equation (3.3) yields:
2V
_
R
x
x
y
+ R
y
y
x
_
IR
x
R
y
=
_
2V +

2
V
x
2
x
2
_
R
y
y
x
+
_
2V +

2
V
y
2
y
2
_
R
x
x
y

IR
x
R
y
xy
=

2
V
x
2
R
y
+

2
V
y
2
R
x
(3.5)
To ease exposition later, instead of working with the current, I(x, y), we work with the current density,
J(x, y) =
I(x,y)
xy
. Setting this into Equation (3.5) yields:
J =

2
V
x
2
1
R
x
+

2
V
y
2
1
R
y
(3.6)
Recalling that the original problem was posed in order to solve the physical problem of supplying energy
accross a microchip, we make the (realistic) assumption that R
s
is uniform on the domain. In this regime:

2
V =

2
V
x
2
+

2
V
y
2
= R
s
J(x, y)
V (x, y) = 0, (x, y)
(3.7)
4 Solutions for Uniform Voltage Requirements
As an intermediate step towards a general solution, we rst consider a special case of the problem that
is, we consider the case where J(x, y) is uniform on the domain. Physically, this corresponds to the setup
in which every transistor in the microchip has the same voltage requirement. This setup can be described
algebraically as:
5
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .

2
V
x
2
+

2
V
y
2
= RJ = constant on (4.1)
V|

= 0 (4.2)
We use a separation of variables V (x, y) = f(x) g(y), expanding each of f(x) and g(y) independently as a
Fourier series. This gives:
V (x, y) =

n=0
_
(f
n
)
1
cos
_
nx
L
_
+ (f
n
)
2
sin
_
nx
L
__

m=0
_
(g
m
)
1
cos
_
mx
L
_
+ (g
m
)
2
sin
_
my
L
__
(4.3)
It follows immediately from the boundary conditions given in Equation (3.7) that (f
n
)
1
= (g
m
)
1
= 0 for all
n, m. Equation (4.3) can therefore be written as:
V (x, y) =

n,m=0
a
n,m
sin
_
nx
L
_
sin
_
my
L
_
(4.4)
Expanding RJ in a similar way:
2
(RJ)
n,m
=
_
2
L
_
2
_
L
0
RJ sin
_
nx
L
_
dx
_
L
0
sin
_
mx
L
_
dy
=
4
L
2
RJ
((1)
n
1)
n
L
((1)
m
1)
m
L
So:
(RJ)
n,m
=
_
0 if n or m is even
16RJ

2
1
nm
otherwise
(4.5)
Substituting the expansion from Equation (4.5) into Equation (3.7) yields:

n,m=0
a
n,m
_
_
n
L
_
2
+
_
m
L
_
2
_
sin
_
nx
L
_
sin
_
my
L
_
=
16RJ

n,m=0
n,m odd
1
nm
sin
_
nx
L
_
sin
_
my
L
_
(4.6)
Equating coecients in Equation (4.6) and using the uniqueness of a Fourier expansion:
a
n,m
=
_
0 if n or m is even
16RJL
2

4
1
nm
1
n
2
+m
2
otherwise
(4.7)
Substituting Equation (4.7) into Equation (4.4) gives the nal expression:
V (x, y) =
16RJL
2

n,m=0
n,m odd
1
nm
1
(n
2
+ m
2
)
sin
_
nx
L
_
sin
_
my
L
_
(4.8)
2
To be absolutely precise, we should say that RJ is only constant and non-zero on

. On , RJ is zero. This allows us to
periodise RJ in an odd function in x and y on [L, L] [L, L]. This periodised function is piecewise continuous, therefore we
expect its Fourier expansion to only converge pointwisely.
6
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
The summation in Equation (4.8) is convergent, so this is the solution of (3.7). That is:

n,m=0
n,m odd
_
_
_
_
1
nm
1
(n
2
+ m
2
)
sin
_
nx
L
_
sin
_
my
L
_
_
_
_
_

n,m=1
1
nm
1
(n
2
+ m
2
)

_
[1,+)
2
1
xy
1
x
2
+ y
2
dxdy

_
[1,+)
2
dxdy
(xy)
2

_
_
[1,+)
dx
x
2
_
2
1 < +
Hence this expansion is normally convergent on , and therefore well dened.
Solving the Problem on our
We can use the symmetry of when seeking the minimum value which V takes on the domain. In particular,
for a J which is constant over the whole domain, the minimum value of V must occur at x = y =
L
2
. Then
Equation (4.8) becomes:
minV = V (
L
2
,
L
2
) =
16RJL
2

n,m=0
n,m odd
1
nm
1
(n
2
+ m
2
)
(4.9)
Improving Convergence
Although Equation (4.9) technically provides a complete solution to the problem for the case when J(x, y) is
uniform on the domain, the evaluation of the right hand side of the expression requires the computation of
an innite series. This innite series converges quite slowly,
3
and so we now formulate a dierent approach
in order to derive an equivalent expression for minV which involves a more quickly convergining sum.
The method works by transforming the Poisson equation into a Laplace equation so that a dierent
separation of variables technique may be employed. In order to ease exposition, we re-parameterise the
domain to be :=
_
(x, y)| x [0, L] , y
_

L
2
,
L
2
_
. Then, setting:
V (x, y) =
RJ
2
x(x L) + (4.10)
it follows that satises:

2
= 0 (4.11)
with the boundary conditions:
(0, y) =(L, y) = 0
(x, L/2) =(x, L/2) =
RJ
2
x(x L).
(4.12)
Equation (4.11) is, therefore, a Laplace equation with nonhomogeneous Dirichlet boundary conditions. We
can then use the method of separation of variables, writing (x, y) = X(x)Y (y), to proceed with the problem.
3
For example, we note that it requires the summation of 19 terms to be able to guarantee accuracy of the answer to just 3
decimal places.
7
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
This yields the following general solution:
(x, y) =

n=0
_
A
n
cosh
_
ny
L
_
+ B
n
sinh
_
ny
L
__
sin
_
nx
L
_
(4.13)
with A
n
and B
n
being two sequences which are yet determined.
Firstly, we note that by symmetry in the geometry of the domain, B
n
= 0. Therefore:
(x, y) =

n=0
A
n
cosh
_
ny
L
_
sin
_
nx
L
_
. (4.14)
Secondly, the second boundary condition given in Equation (4.12) implies that:
(x, L/2) = (x, L/2) =

n=0
A
n
cosh
_
n
L
L
2
_
sin
_
nx
L
_
=
RJ
2
x(x L). (4.15)
Now, the last term in Equation (4.15) has the following Fourier expansion:
RJ
2
x(x L) =
2RJ L
2

n=1
(1)
n
1
n
3
sin
_
nx
L
_
. (4.16)
By the uniqueness of the Fourier expansion, we have:
(x, y) =
4RJ L
2

n=0
1
(2n + 1)
3
cosh
_
(2n+1)
L
y
_
cosh
_
(2n+1)
2
_ sin
_
(2n + 1)x
L
_
(4.17)
It is straightforward to show that the series on the right hand side of Equation (4.17) is absolutely convergent,
and furthermore the cosh terms in the expression guarantee very rapid convergence of the sum (as desired).
is, therefore, well-dened, and (again by symmetry) under this parameterisation we note that the maximum
voltage drop is attained at the postion (L/2, 0). Its exact value is given by:
V (L/2, 0) = RJ L
2
_
_
1
8

4

n=0
1
(2n + 1)
3
(1)
n
cosh
_
(2n+1)
2
_
_
_
. (4.18)
Relating Back to Industry
At this point, the special case of uniform voltage requirements across the domain is solved. Using a computer
to accurately compute the value of the sum to a large number of decimal places, we are able to report back
to the industrial experts that in this setup the maximum voltage drop occurs in the middle of the domain,
and has value:
max
(x,y)
V (x, y) =
16R
s
JL
2

4
0.448516 . . . = 0.0736 . . . R
s
JL
2
Note that using Equation (4.18), using only the rst term of the expansion, one nds the same value with the
same number of decimal places (0.0736). We numerically solved Equation (3.7) with a nite element solver,
and we plot its solution in Figure (3). We note that the maximum value for V found numerically corresponds
well with the theoretical one.
8
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
Figure 3: Numerical Solution of Equation (3.7)
5 Single Block Current Density
As another intermediate step towards a general solution, we now consider the case where only a small
rectangular subset of the domain has a current requirement, and where the current density is zero elsewhere.
Intuitively, a similar approach as for the uniform current density can be applied, taking into account the
new boundaries of the inner block corresponding to the non-zero current density. For simplicity, we will keep
the notation J for the current density in the inner block (which we henceforth denote as B = [x
1
, x
2
][y
1
, y
2
]).
Such a system is governed by the equation:

2
V
x
2
+

2
V
y
2
= RJ
B
(x, y)
V = 0 on ,
(5.1)
where
B
(x, y) is the characteristic function for the inner block B.
The solution of Equation (5.1) has the Fourier expansion solution:
V (x, y) =

n,m=0
a
n,m
sin
_
nx
L
_
sin
_
my
L
_
. (5.2)
Again, we decompose the current density via the Fourrier transfom:
RJ =

n,m=0
RJ
n,m
sin
_
nx
L
_
sin
_
my
L
_
, (5.3)
The Fourier coecients (RJ)
n,m
are given by:
(RJ)
n,m
=
4
L
2
_
x
2
x
1
_
y
2
y
1
RJ sin
_
nx
L
_
sin
_
my
L
_
dxdy
=
16

2
1
n m
sin
_
nl
x
2L
_
sin
_
nc
x
2L
_
sin
_
nl
y
2L
_
sin
_
nc
y
2L
_
,
(5.4)
where l
x
and c
x
represent the length and center of the x boundary interval of the inner block respectively,
9
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
Figure 4: Numerical Solution of Equation (5.1) with a non-zero current density in [0.3, 0.4] [0.3, 0.4]
and l
y
and c
y
are dened similarly. By the uniqueness of the Fourier expansion, we have:
a
n,m
=
16RJL
2

4
1
n m
1
n
2
+ m
2
sin
_
nl
x
2L
_
sin
_
nc
x
2L
_
sin
_
ml
y
2L
_
sin
_
mc
y
2L
_
. (5.5)
Noting that
1
n
2
+m
2

1
nm
, it follows that |a
n,m
|
16RL
2
J

4
1
n
2
m
2
. We therefore obtain the bound:
|V (x, y)|
4
9
RJL
2
x, y (5.6)
As a check on our analytical results, we solve Equation(5.1) numerically using a nite element solver. The
results are displayed in Figure 4. Crucially, the maximum voltage drop is indeed below the bound given by
Equation (5.6).
6 Multiple Block Current Density
In Sections 4 and 5, we have dealt with simple cases of the general problem, to which we turn our attention
now. We note that the solution to Equation (5.1) can be generalised to the case where the domain is
not merely comprised of a uniform or isolated region with uniform voltage requriements, but is instead
partitioned into a collection of rectangular blocks (B
i
). More precisely, we now consider the general case
where =

i
B
i
, with B
i
B
j
= i, j, such that region B
i
as current density J
i
and resistivity R
i
.
Extending the simple case explored in Section 5, we note that the system is now governed by the equation:

2
V =

i
R
i
J
i

B
i
(x, y)
V = 0 on
(6.1)
By the superposition principle [2] and linearity of Equation (6.1), the general solution is simply a sum of
solutions of the type derived for Equation (5.1) with non-zero current densities of dimensions (l
x
i
, l
y
i
) found
10
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
at (c
x
i
, c
y
i
):
V (x, y) =
16L
2

n,m=0
_
a
n,m
1
n m
1
n
2
+ m
2
sin
_
ny
L
_
sin
_
my
L
_
_
with a
n,m
=

i
R
i
J
i
sin
_
nl
x
i
2L
_
sin
_
nc
x
i
2L
_
sin
_
ml
y
i
2L
_
sin
_
mc
y
i
2L
_
(6.2)
7 Optimal Chip Congurations
Our results from Sections 4, 5 and 6 now allow us to advise the industrial experts on the optimal conguration
of a given set of blocks (B
i
) on a given microchip i.e., the conguration which yields the lowest maximum
global voltage drop among all possible (and, geometrically speaking, admissible) congurations .
If a chip is composed of p
2
blocks of the same dimension l, there are (p
2
)! ways to arrange them on .
The optimal conguration,
opt
((c
x
1
, c
y
1
), (c
x
2
, c
y
2
), . . . , (c
x
p
2
, c
y
p
2
)) satises:

opt
((c
x
1
, c
y
1
), (c
x
2
, c
y
2
), . . . , (c
x
p
2
, c
y
p
2
)) = arg min

max
x,y
V

(x, y) (7.1)
Developing an algorithm for the placement of blocks on is, in general, an NP complete problem [1].
Although we are unable to oer such an algorithm explicitly, we are able to oer observations which govern
the optimal solution. In the case where all blocks have the same resistivity:
1. If there are 4 or fewer blocks to place on the chip, they should each be placed in a corner of the chip.
2. If there are more than 4 blocks to place on the chip, the 4 with the largest voltage requirements should
be placed in the 4 corners of the chip.
3. All blocks should be placed as far away from each other as possible.
To illustrate some of these ideas, we have included in the appendix some gures of the numerical solutions
of Equation (6.1), with blocks placed in certain congurations. Figure 5 shows the dierence between placing
a single block in the centre or at the edge of . Figure 6 shows the dience between placing blocks close
to or far from each other. Figure 7 shows the dience between placing those blocks with largest voltage
requirements between or outside those blocks with lower voltage requirements.
8 Conclusions
By making use of a Fourier expansion to solve a Poisson equation on a rectangular domain, we initially
solved some simple cases of the problem posed. We then explored how these simple cases were actually
the building-blocks for the mechanics of the general problem that is, the optimal placement of multiple
blocks (with dierent voltage requirements) on the domain. Our general solution takes the form of a Fourier
expansion, and we make use of the superposition principle and of linearity of the governing equation in order
to generalise the simple cases to the main problem.
Our results have allowed us to develop an intuition for the optimal conguration of the microchip, and
have been veried via the use of a nite element solver. Our modelling assumptions have been moderate
throughout, but it should be noted that stringent testing of the physical setup would be required before
our models could be adopted by chip manufacturers in industry. In particular, our assumption of uniform
sheet resistance in the wire mesh may prove to be unrealistic due to the extremely small scales involved in
manufacture minor imperfections of the orders of microns would almost certainly cause this assumption to
11
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
be violated. Also, we have neglected to factor in other variables such as temperature, which aects resistance
in the mesh and is of reasonable concern in a microchip (which is known to get hot during operation).
This being said, we believe our work to be instructive of the general and most important behaviours
governing the voltage drop on the microchip, and think that our work could be easily modied if it were the
case that these external factors were proven to be more inuential than we have taken account for.
9 Acknowledgements
Our sincerest thanks go to:
Maria Aguareles for mentoring us
Kate Lewin and Chris Breward for organising the OCCAM Graduate Modelling Camp event
OCCAM and KAUST for their generous funding
References
[1] M. Garey and D. Johnson, Computers and Intractability: A Guide to the Theory of NP-Completeness,
WH Freeman and Co Ltd, 1979.
[2] K. Riley, M. Hobson, and S. Bence, Mathematical Methods for Physics and Engineering, 3rd Ed: A
Comprehensive Guide, Cambridge University Press, 2006.
[3] P. Tipler and G. Mosca, Physics for Scientists and Engineers, WH Freeman and Co Ltd, 2007.
[4] P. V. Zant, Microchip Fabrication, 5th Ed: A Practical Guide to Semiconductor Processing, McGraw-
Hill, 2004.
12
OCCAM Modelling Camp Analysis of the Static Voltage Drop. . .
Appendices
Figure 5: Solution of 6.1 with a single block, placed either at the edge (left) or in the middle of (right)
Figure 6: Solution of 6.1 with a a pair of blocks, placed far from (left) or close to (right) each other
Figure 7: Solution of 6.1 with the blocks with highest voltage requirements placed towards the edge (left) or
towards the middle of (right), and the other blocks lling in the gaps
13

Você também pode gostar