( ESNUG 486 Item 9 ) -------------------------------------------- [10/08/10]
From: Chris Van Meter <cva=user domain=cypress got calm>
Subject: Benchmarking new Calibre SmartFill vs. old Calibre dummy fill
Hi, John,
We've used Calibre's SmartFill for partial mask and full mask tapeouts.
SmartFill is Calibre's new dummy metal fill functionality. It optimizes
the quantity, shape and location of metal fill to reduce CMP (chemical-
mechanical polishing) variability, or smoothing of the cross-die metal
layers.
Before using SmartFill, we used Calibre's generic algorithm to place
'dummy fill' metal on our design's open regions to the highest allowable
density. The downside of maximizing or blindly placing dummy fill was
that our design could become overfilled, ending up with unaccounted
parasitic capacitance from the extra metal, which can negatively affect
timing. We had concerns about this, so our goals were to:
- Add as little fill as possible and still meet our metal density
constraints and avoid unwanted electrical effects.
- Achieve a wafer planarization (highest density to lowest density)
that was as smooth and devoid of large density variations as
possible. This decreases polishing time and can potentially lead
to improvement in yield.
A while back, we decided we wanted to test and implement intelligent dummy
fill. Although we hadn't had any actual failures due to dummy fill, we felt
that having the power to manipulate dummy fill generation would allow us to
achieve lower density in designs.
Calibre's SmartFill functionality for DFM appealed to us because we already
used its SVRF (Standard Verification Rule Format) for physical verification.
Integrating this Yield Enhancer feature was minimal. We simply modified our
deck using a different SVRF command. It just required additional Calibre
Yield Enhancer licenses. After thoroughly beta testing our old Calibre
dummy fill approach versus SmartFill, we moved to incorporating it as part
of our process.
The SmartFill feature has constraints for:
- minimum density
- maximum density
- magnitude density (minimizing difference between highest
and lowest density across the die)
- gradient density (minimizing difference between highest
and lowest density between adjacent windows)
SmartFill conducts a density analysis, which it checks against the original
density of the design. SmartFill adds the fill shapes and attempts to add
as little fill as possible to meet all the density constraints set by the
user. For instance, if the design cannot meet a magnitude constraint in its
current state because one window has high density, it will attempt to raise
density in the lowest window by adding dummy fill to meet the constraint.
Calibre SmartFill vs. old Calibre dummy fill:
Using the old Calibre to place dummy fill was time consuming and tedious.
We typically had a handful of layers and generally went through multiple
dummy fill passes for each layer. Old Calibre was semi-automated and would
require some design time. Getting our density to acceptable levels took
approximately 5 man-days.
SmartFill is very user friendly -- just switch on the "DFM SpecFill" SVRF
command to turn on the SmartFill algorithm. Most full chip runtimes are
around 2-3 hours for all dummy fill layers generated with a SmartFill run.
In contrast to the old Calibre dummy fill, SmartFill was single pass and
encompassed all layers in one run.
On average, we saved about 4 man-days of total time for metal fill vs.
doing it with a two-pass algorithm with Calibre's RECTANGLES command alone.
Plus every time we taped out the mask, we had to redo the metal fill, which
results in even greater time savings due to SmartFill's automation.
Our typical time savings with the new SmartFill was around 2 man-weeks.
Gotchas we found with SmartFill:
- When we first started evaluating SmartFill, we still had to do some
CAD manipulation by inserting dummy values to force the tool to meet
certain constraint parameters. There were a few test cases where
the range constraint was failing on the low end. In order to get
the SmartFill to meet constraints, we added lower limit target values
to add as criteria for the tool as well. Mentor fixed SmartFill to
resolve these problems, and we have been removing the unnecessary
coding in our decks that we had used to handicap the older versions.
- We have experienced one area of failure with SmartFill. The process
for adding dummy fill is isolated to a single die or frame. The
breakdown occurs when doing experimental tapeouts with test die
that are not as dense as typical product dies. What occurs is the
test die has lower natural density, so the tool adds little fill to
meet the density constraints. Then a real product die which has
original higher density undergoes dummy fill to meet density
constraints as well. When these two separate entities, product die
and test die, are placed in a frame or reticle adjacent to another
there is a potential gradient problem for densities. These in turn
cause failures in processing. On a wide scale, adding dummy fill at
the top level die & frame would be the solution. Two drawbacks to
this process: First, mask sets for dies and frames are delivered
separately. The other drawback is at this time in the design
process it would be a nightmare to resolve density problems inside a
design and remain on schedule.
Benchmarking SmartFill:
Below is an example of reduced fill shapes on a recent design. Notice how
SmartFill reduced our overall polygon count, while at the same time getting
our required pattern density. In the table below, the pattern density
percentages = metal or diffusion area/window area.
Polygon Count and Pattern Density Comparison for the
old Calibre Dummy Fill vs. the new Calibre SmartFill
Min Max Range Shapes Notes
FOM layer
Calibre Rect 38.0% 65.4% 27.4% 9,519,600
SmartFill 35.5% 63.0% 27.6% 1,726,000
----- ----- ----- ---------
%Diff 7.0% 3.8% -0.7% -82.1% Big drop in dummy fill!
P1M layer
Calibre Rect - - - 3,168,708
SmartFill 22.8% 58.3% 35.5% 606,520
----- ----- ----- ---------
%Diff - - - -80.6% Big drop in dummy fill!
SmartFill is part of Cypress' Calibre Created Layer DRC flow and a standard
part of our physical verification process. We have been very pleased with
the tools performance and results. Mentor Graphics continues to have
exceptional support for it both at the factory and on-site.
- Chris Van Meter
Cypress Semiconductor Corp. Lexington, KY
Join
Index
Next->Item
|
|
