Scribd
Upload
76 views

CS4617 Computer Architecture: Lecture 2a

This document discusses computer architecture lecture notes covering topics like die yield, yield calculation formulas, costs associated with fabrication processes, Amdahl's law, and speedup calculations for parallel programs using multiple processors. Equations show how yield is affected by die area, defect density, and process complexity, and how performance scales with numbers of parallel processors depending on parallelizable code fractions and communication overhead.

Uploaded by

dzakybd
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
76 views

CS4617 Computer Architecture: Lecture 2a

This document discusses computer architecture lecture notes covering topics like die yield, yield calculation formulas, costs associated with fabrication processes, Amdahl's law, and speedup calculations for parallel programs using multiple processors. Equations show how yield is affected by die area, defect density, and process complexity, and how performance scales with numbers of parallel processors depending on parallelizable code fractions and communication overhead.

Uploaded by

dzakybd
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 13

CS4617 Computer Architecture

Lecture 2a

Dr J Vaughan

September 15, 2014

1/13
Die yield

I Fraction of good dies on wafer = die yield


I Die yield =
Wafer yield × 1/(1 + Defects per unit area × Die area)N
I This is the Bose-Einstein formula: an empirical model
I Wafer yield accounts for wafers that are completely bad, with
no need for testing
I Defects per unit area accounts for random manufacturing
defects = 0.016 to 0.057 per cm2
I N = process complexity factor, measures manufacturing
difficulty
= 11.5 to 15.5 for a 40nm process (in 2010)

2/13
Yield

I Example
I Find the number of dies per 300mm wafer for a die that is 1.5
cm square.
I Solution
Die area = 1.5 × 1.5 = 2.25cm2
π × (30/2)2 π × 30
Dies per wafer = −√
2.25 2 × 2.25
= 270

3/13
Yield

I Example
I Assuming the wafer yield is 100%, find the die yield for a die
that is 1.5 cm square
I The defect density is 0.031 per cm2 and the process
complexity factor is 13.5
I Solution
1
Die yield =
(1 + 0.031 × 2.25)13.5
= 0.4

I Note
I Defect density per chip = 0.031 × 2.25 = .06975 ≈ 7%
I Why is yield not 93%?

4/13
Other costs

I The computer designer affects die size and cost by


including/excluding functions and by the number of I/O pins.
I Mask costs for a high-density fabrication process with 4–6
layers exceed $1M
I This is a fixed cost that is likely to increase
I It makes small-volume production very expensive

5/13
Problem

I What power savings can be made by cutting processor voltage


by 20% and clock frequency by 40%?
(Voltage×0.80)2
I Energynew = = 0.82 = 0.64
Energyold Voltage 2
switched ×0.6
I Powernew
Powerold = 0.64 × Frequency
Frequencyswitched = 0.384

6/13
Problem

I An enhancement to a computer accelerates some mode of


execution by a factor of 10. The enhanced mode is used 50%
of the time, measured as a percentage of the execution time
when the enhanced mode is in use. Recall that Amdahl’s Law
depends on the fraction of the original unenhanced execution
time that could make use of enhanced mode. Thus the 50%
measurement cannot be used directly to compute speedup
with Amdahl’s Law.
I What is the speedup obtained from fast mode?
I What percentage of the original execution time has been
converted to fast mode?

7/13
Approach to solution

I Up to now, we have phrased Amdahl’s Law in terms of the


fraction of an “old” system that can be enhanced. Call this
Fenhanced
I This time, we are given the fraction of execution time that the
enhanced component is used after the enhancement has taken
place. Call this Fdehanced to show that we can reverse the
process to convert it to the original fraction that can be
enhanced
I First, apply Amdahl’s Law in the normal “forward” direction
to get an expression for the ratio of new to old execution times
I Then, apply Amdahl’s Law backwards in time to get an
expression for the ratio of old to new execution times

8/13
Solution

Exectimenew 1
= (1 − Fenhanced ) + Fenhanced ×
Exectimeold speedupenhanced
= (1 − Fenhanced ) + 0.1 × Fenhanced
= 1 − 0.9 × Fenhanced
Exectimeold 1
= (1 − Fdehanced ) + Fdehanced ×
Exectimenew speedupdehanced
0.5
= 0.5 +
0.1
= 5.5
1
5.5 =
1 − 0.9 × Fenhanced
5.5 − 4.95 × Fenhanced = 1
4.5
Fenhanced =
4.95
= 0.91
9/13
Check

1
Speedup = 1
(1 − 0.91) + 0.91 × 10
1
=
0.09 + 0.091
1
=
0.181
= 5.52

10/13
Problem

I What is the speedup with N processors if 80% of the


application is parallelizable, ignoring the cost of
communication?
I

Exectimeold 1
= 1
Exectimenew (1 − Fenhanced ) + Fenhanced × speedupenhanced
1
= 0.8
0.2 + N

11/13
Problem

I What is the speedup with 8 processors if, for every processor


added, the communication overhead is 0.5% of the original
execution time?
I N=8
Exectimeold 1
= 1
Exectimenew (1 − Fenhanced ) + Fenhanced × speedupenhanced
1
= 0.8
0.2 + 8 + 0.005 × 8
1
=
0.34
= 2.94

12/13
Problem

I What is the speedup with 8 processors if, for every time the
number of processors is doubled, the communication overhead
is increased by 0.5% of the original execution time?
I N=8
Exectimeold 1
= 0.8
Exectimenew 0.2 + 8 + 0.005 × 3
1
=
0.315
= 3.17

13/13

You might also like