Title: Reticle Floorplanning With Guaranteed Yield for MultiProject Wafers
1Reticle Floorplanning With Guaranteed Yield for
Multi-Project Wafers
Andrew B. Kahng
Sherief Reda
ECE and CSE Dept. University of California San
Diego
CSE Dept. University of California San Diego
2- Introduction to Multi-Project Wafers
- Design Flow
- Side-to-Side Dicing Problem
- Proposed Methodology
- Floorplanning with Guaranteed Yield
- Experimental Results
3- Mask cost 1M for 90 nm technology
- Wafer cost 4K per wafer
4Image courtesy of CMP and EuroPractice
- Share rising costs of mask tooling among
multiple prototype and low production volume
designs ? Multi-Project Wafer
5- Introduced in late 1970s and early 1980s
- Companies MOSIS, CMP, TSMC
- Several academic approaches proposed
- Chen et al. give bottom-left fill algorithm, SPIE
2003 - Xu, Tian, Wong and Reich, SPIE 2003
- Anderson et al. propose a grid packing algorithm,
WADS 2003 - Kahng et al. propose dicing plans for generic
floorplans, ISPD 2004
- Commercial Tools MaskCompose, GTMuch
6- Introduction to Multi-Project Wafers
- Design Flow
- Side-to-Side Dicing Problem
- Proposed Methodology
- Floorplanning with Guaranteed Yield
- Experimental Results
7 8- Custom designs
- Partition between reticles
9- Custom designs
- Partition between shuttles
- Reticle placement
10- Custom designs
- Partition between shuttles
- Reticle placement
- Stepper shot-map
shot-map
print
11- Custom designs
- Partition between shuttles
- Reticle placement
- Stepper shot-map
- Dicing plan design
12- Custom designs
- Partition between shuttles
- Reticle placement
- Stepper shot-map
- Dicing plan design
13- Custom designs
- Partition between shuttles
- Reticle placement
- Stepper shot-map
- Dicing plan design
- Extract die
14- Introduction to Multi-Project Wafers
- Design Flow
- Side-to-Side Dicing Problem
- Proposed Methodology
- Floorplanning with Guaranteed Yield
- Experimental Results
15- Side-to-side dicing is the prevalent wafer
dicing technology
Sliced out
- A die is sliced out if and only if
- 1. Four edges are on the cut lines
- 2. No cut lines pass through the die
Dicing is easy for standard wafers. All dice will
be sliced out.
Dicing is complex for MPW. Most dice will be
destroyed if placement is not well aligned.
16- Two dies are in conflict if they can not be
simultaneously sliced out horizontally. - Die 1 is in conflict with entire row of Die 2.
17dicing die 2
1
1
2
2
3
4
3
4
6
6
5
6
6
dicing die 2
die
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
3
4
3
4
3
4
3
4
3
4
3
4
3
4
3
4
6
6
5
6
6
5
6
6
5
6
6
5
6
6
6
6
6
6
6
6
dicing die 1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
3
4
3
4
3
4
3
4
3
4
3
4
3
4
3
4
6
6
5
6
6
5
6
6
5
6
6
5
6
6
6
6
6
6
6
6
dicing die 1
1
1
2
2
wafer
3
4
3
4
reticle
6
6
5
6
6
- A die copy is successfully extracted if it is
successfully extracted in both horizontal and
vertical dicing
18Given
A number of die designs, each with a production
volume requirement
Objective
Produce the required volumes of all dies using
the minimum amount of wafers ? maximize the
minimium amount of die copies extracted from a
wafer over all dies (Yield)
Example If die 1 has a 40 copies requirement,
and the dicing plan yields 5 valid copies per
wafer ? 8 wafers are needed
Methods
- Floorplanning the dies within the reticle
- Efficient dicing plan
19- Introduction to Multi-Project Wafers
- Design Flow
- Side-to-Side Dicing Problem and Motivation
- Proposed Methodology
- Floorplanning with Guaranteed Yield
- Experimental Results
20- Previous approaches (Kahng et al.) use regular
area floorplanning and devise efficient dicing
plans
Observation
- Regular area floorplanners are yield oblivious
- There is an inherent conflict between area
minimization and yield
21Main Idea
- Construct floorplans that consider specified
yield bounds as constraints and minimize the area
Method
- Limit floorplans to grids
- Calculate constructive lower bounds on the yield
- Specify simple rules to characterize the yield of
a given floorplan - Minimize the area using a branch and bround
procedure taking the yield as a constraint
22General floorplan
Grid floorplan
1
1
2
2
3
4
3
4
5
6
6
6
6
23n
- Given m reticle rows and n reticle columns
reticle row
- Let ?i be the number of different height die
copies residing in a grid row, e.g., ?12, ?11 - Let µj be the number of different width die
copies residing in a grid column, e.g., µ11,
µ22
grid row
m
reticle column
grid column
Lemma 1 A constructive lower bound on number of
copies that can be extracted of a die residing in
row i and column j is ?m / ? i? ? ?n / µj?
24- ?i number of different height die copies
residing in a grid row. - µj be the number of different width die copies
residing in a grid column
reticle row
grid row
- Limiting the values of µj and ?i guarantees a
lower limit on yield
reticle column
grid column
Example
Theorem 1 A lower bound on the yield is
m2-m(?i?µj)/?i?µj for a wafer with m rows and m
columns
25- Why is branch and bound feasible?
- 1. There are typically few dies per reticle
- 2. Yield constraints prune large portions of the
search space
- branch_bound(k, yield, grid1..w1..h)
- if all dies are placed then
- evaluate the floorplan area and if area lt best
area then area best area - return
- expand the grid by an additional column and row
grid1..w11..h1 - for each empty slot (i, j) in the grid
- if placing die dk in slot (i, j) does not
violate yield constraints - Place dk and evaluate partial area
- if partial area lt best area then
branch_bound(k1, yield, grid1..w11..h1) - undo placement of dk in (i, j)
26- Introduction to Multi-Project Wafers
- Design Flow
- Side-to-Side Dicing Problem and Motivation
- Proposed Methodology
- Floorplanning with Guaranteed Yield
- Experimental Results
27- Our method allows an area/yield trade off
28- Our results dominate previous approaches in both
yield and area
29- Our results establish a Pareto frontier
representing a trade off between area and yield
30Conclusions
- A new simple approach to reticle floorplanning
taking yield as a constraint - An optimal area packer using branch and bound
- Our results establish a Pareto frontier that
trades yield for area - Our results dominate previous approaches
Future Work
- Respecting reticle aspect ratios
- Different dicing plans for different wafers
31(No Transcript)