GDSII-based Flow Speeds Mask Data
Preparation
With the
65-nanometer process technology node on the horizon, pressure is growing
in the data preparation and mask manufacturing communities to create ways
to handle an increasingly complex data flow. The source of exploding data
volumes is several fold. The advancements inspired by Moore's Law have
resulted in complex systems-on-chip.
More and more aggressive resolution enhancement techniques (RET), such
as optical process correction (OPC), phase shift masks (PSM) and
scattering bars are being developed and deployed in order to achieve the
required feature sizes. Complex pattern treatments and the splitting of
single mask layers into multiple masks, such as in the case of phase
shifting masks, have resulted in increased data volumes.
The International Technology Roadmap for Semiconductors (ITRS) predicts
file sizes as large as 150 Gbytes in 2004. The result is not only a
problem of logistics for storage and data transfer, but also a problem of
longer processing times that significantly impact throughput and
turn-around time (TAT).
Another important contributor to exploding data volumes involves the
evolution of the mask writing process. Mask data preparation used to be a
commodity mostly done by the tapeout group affiliated with the design
house. The process was fairly simple: MEBES and similar derived formats
became the common exchange format.
Instead of a design database hand-off, flat mask layer-specific files
were transferred. This was perceived to give some measure of security,
especially since data was often sent to the mask manufacturer in the
physical form of a magnetic tape. In this process, the mask house did not
treat the data unless necessary. But as the technology roadmap advanced,
things changed.
Technology advancements contribute to data volumes
In the 1990s, raster-scan laser mask lithography tools were introduced.
These tools were designed to be compatible with the existing MEBES mask
data; however, the high contrast optical resists used with these tools
required sizing to get the best critical dimension (CD) performance.
Write strategies were changed to deal with throughput issues. For
instance, data was often "smashed" to improve registration between
multiple patterns on the same mask. Smashing, which merges fractured
pattern files, then re-fractures the files again, results in the mask
supplier having to make routine mask data modifications; that is, the mask
supplier starts re-fracturing the data.
The sizing operation and smashing of a predefined job deck, as well as
reverse toning, are considered simple operations that have a minimal
chance of changing the content of the data. This made the process
acceptable to the customer. But at the same time, the data communication
structure started changing. Magnetic tape gave way to networks as a
popular and easy means for data exchange. Storage needs grew in order to
accommodate not only the transferred file, but also the intermediate
results of the re-fracture step and the optimized job decks.
Adding to the burden of data containment were shrinking geometries and
new technology advances. Dry-etching, optical mask writing machines,
proximity correction steps, and variable shaped beam (VSB) mask writing
machines all impacted data volumes, many requiring specific formats for
the mask writing process. Many of the formats had different architectures
that were not compatible with the MEBES format.
For instance, the MEBES format is a flat format; the VSB formats
support hierarchy to different degrees. (The flow created by these
processing options is shown in figure 1.) The target format of the initial
flow was transformed to an exchange format between point tools, which are
normally concatenated by an automation scheme. The data set now passes
through a number of data preparation stations. A large number of files
need to be archived to document the process, depending on the path of the
individual job.
Figure 1 — Schematic mask data flow for advanced
e-beam and laser mask writers including support for mask manufacturing
process enhancements.
This complex flow is the current operation model for a number of
businesses. It allows maximum flexibility in tool allocation and loading,
and an exchange between multiple manufacturing sites with little
interruption for the manufacturing process, but it has two major drawbacks
that result in serious consequences for nanometer manufacturing.
First, the overhead in file transfer and loading has increased
dramatically; its impact on hardware and turn-around-time cannot be
neglected. Files must be handled in the chosen exchange format. This
precludes the use of the hierarchy in the original design data, which
would reduce the file size and deploy optimized data preparation
algorithms.
Second, a re-fracture step is commonly used to conduct the geometry
processing steps. Figure 2 shows the average runtime distribution for
three functions in a data preparation run: Boolean layer operation, sizing
and fracturing.
Figure 2 — Typical runtime distribution for data
preparation jobs targeting MEBES and JEOL output formats. The process
starts with a hierarchical GDS-II file and includes layer combination and
sizing alongside with typical mask transformations.
Sizing and Boolean operations require about 10% of the total processing
time, whereas fracturing takes 80% of the time. With any resizing step
beyond the original fracture step that creates an 80% overhead, it can be
assumed that the data has been supplied in a fractured state. This flat
exchange format prohibits the leveraging of hierarchical fracture methods
that allow for faster throughput.
The evolution of semiconductor manufacturing has enabled more
functional chip design, but the current standard flow has created a
crippling data bottleneck. An analysis of this standard flow leads to a
number of conclusions:
-
The data is fractured too early in the flow.
-
Hierarchy is destroyed too early in the process.
-
The current commonly used exchange format is geared too much towards
MEBES compatible writing tools and does not address other machines in
the flow sufficiently.
-
Re-processing of data is an unnecessary overhead.
But how
does a mask shop halt the problems of a data bottleneck without massive
investment in hardware or a disruptive change to the system? A practical
and simple solution is to use an open hierarchical format for data
exchange and transfer.
Streamlining with a GDS-based flow
A GDS-based flow enables an efficient exchange of data between the
point solutions in the current flow. Depending on the path through the
flow, the formatting/fracturing step is conducted at the last moment, just
prior to mask writing. Introduction of biases, tool switches and
manufacturing site transfer related changes to the data are efficiently
conducted in a hierarchical format and use about 10 - 20% of the time it
currently requires. The open format also enables the interoperation of
different tools in the same flow and an easy extension into new functions
if required.
Hierarchy preservation and management in the data flow has many
benefits. Geometry count (a measure of the file size) is reduced
significantly when hierarchy is present. An efficient hierarchical engine
optimizes the hierarchy during the read-in step and reduces the geometry
count yet further.
The benefit of hierarchy pays off particularly well if a sufficient
number of placements are present, specifically in large arrays often found
in memory chips. The development trend in microprocessors shows a growing
portion of the area is hierarchy-aware; flow optimization allows hierarchy
to be maintained even in complex shape processing. The same holds for
systems on chip. Figure 3 shows examples of the hierarchy benefit for the
file size after the application of model-based OPC.
Figure 3 — Hierarchy in the database after the
application of OPC. For comparison the output files have been
intentionally flattened. The experiment was conducted on subset of the
original file and then averaged for each layer.
Even as ground rules shrink, efficient OPC can preserve a significant
portion of the hierarchy. Figure 4 evaluates the data ratio for flat
versus hierarchical OPC depending on ground rule and optical diameter. It
illustrates that hierarchy utilization and preservation will still yield a
5-fold benefit at 65nm, even with an aggressive optical diameter.
Figure 4 — Data volume ratio for flat vs.
hierarchical OPC for various ground rules depending on the optical
diameter for a poly layer. An efficient OPC tool can retain and utilize
hierarchy even as the ground rules shrink.
Test results prove GDS hierarchy improves flow
Experiments conducted at a prominent mask house illustrate the benefit
of hierarchy in the mask data preparation flow. Portions of various test
cases were fractured hierarchically starting with a GDS-II file. The run
times and file sizes were recorded. The same portions were then flattened
completely and processed again. Figure 5 shows the processing time ratios
and for the MEBES fracture. Files were written into MODE 5 with 64-stripe
compaction, and typical orientation commands were conducted during the
run.
Figure 5 — MEBES fracture time comparison for
hierarchical and flat input databases. The experiment simulates the
situation of a re-fracture step of a flat database vs. a hierarchical
input file. The flattening was conducted as an independent step prior to
the processing.
Similar tests were conducted on the JEOL fracture format (Fig. 6). A
field size of 500m was chosen, small figures were suppressed, and typical
mask orientation steps were conducted. The data proved a significant
benefit in throughput for the formatting step that started with a
hierarchical database compared to a fracturing step conducted on a flat
database. In addition, the resultant file sizes are smaller when hierarchy
is present, since the JEOL format supports hierarchy itself. The benefit
of the latter outweighs the overhead induced by the hierarchical fracture
approach.
Figure 6 — JEOL fracture time comparison for
hierarchical and flat input databases. The experiment simulates the
situation of a re-fracture step of a flat database vs. a hierarchical
input file. The flattening was conducted as an independent step prior to
the processing.
Figure 7 shows a case study of the processing time benefit for the
following flow:
The
assumption in the first case is that a fractured flat file enters the
process, whereas in the second case, a hierarchical GDSII file is passed
in. Geometry processing steps are valued 2 processing time equivalents and
fracturing processing steps are valued 8 processing time equivalents (a
20:80 time ratio). The comparison shows the potential of a 2.8X throughput
benefit even if the individual tool performance is assumed to be
equivalent.
Figure 7 — Case study on TAT improvement by
introducing a hierarchical exchange format and eliminating fracturing for
geometry processing for an advanced mask flow that includes mask OPC and
density dependent bias.
Looking ahead
So far the benefits of a hierarchical format and the hierarchical
processing capability have illustrated its capability to reduce the TAT of
various data processing steps. Data volume reductions are possible through
the reduction of the number of representations of the same data. The flow
can be enhanced by integrating geometry-handling steps into a single run,
thus eliminating intermediate files.
Duplication of files can also be omitted if all tools in the flow can
read and write the common format and no translations are necessary. A
close comparison of the file sizes for individual layers shows that the
mask writer formats can be more efficient in representing the data than
GDS-II. Figure 8 compares the file sizes in different formats for two
layers — there is clearly a need for a more efficient representation of
design data.
Figure 8 — File sizes for two test cases in
different formats. An experimental format was used for the GDS-II
replacement experiment.
Analysis of alternate experimental file formats suggests that
significant file size reduction compared to GDS-II is possible. Figure 8
includes experimental data from the development of a GDSII replacement
called OASIS. It has shown to complement the benefit of turn around time
with a reduction in storage need and time for data exchange. File size
improvements up to a factor of 5-50 have been demonstrated on a broader
range of test cases.
An alternate mask data preparation flow, based on GDS-II or its more
efficient replacement, OASIS, offers a solution to the challenges of
growing file sizes and increased turn-around times. The new flow targets
the elimination of re-fracturing steps for a flexible data dis-positioning
between tools and manufacturing sites.
The introduction of an open hierarchical format enables the leverage of
hierarchical shape processing throughout the flow. The flow efficiency
gained by this approach will reduce the total volume of data moved and
stored in the process. A more efficient hierarchical format has the
potential to significantly reduce the data volume beyond that
significantly.
The industry standard database is GDSII; it is the database that all
other databases, including proprietary ones, must stream-out to prior to
handing off to the manufacturer. Selecting tools that operate in GDSII
originally not only saves an error-prone data transfer stage, it also
ensures that data retains hierarchy, thereby containing file size, and
containing it in the correct format for manufacturing. For mask writing,
it is particularly important to contain hierarchy as it helps manage file
size, streamlines the mask writing flow, and enhances turn around time.
Figure 9 — Current and suggested delivery
flow.
Steffen Schulze, Ph.D., is a product marketing manager for Mentor
Graphics, working with the Calibre mask data preparation tool suite.
Pat Lacour is a technical marketing engineer for Mentor Graphics,
specializing in the Calibre resolution enhancement technology product
line.
Peter Buck is a senior member of the technical staff at DuPont
Photomasks, Inc.
