Simplify to Survive,

prescriptive layouts ensure profitable scaling to 32nm and beyond

Lars Liebmann1, Larry Pileggi2, Jason Hibbeler1, Vyacheslav Rovner2,
Tejas Jhaveri2, Greg Northrop1
1
IBM Microelectronics, 2070 Route 52, Hopewell Junction, NY 12533
2
PDF Solutions, Pittsburgh 5830 Ellsworth Ave, Suite 304. Pittsburgh PA 15232

ABSTRACT

The time-to-market driven need to maintain concurrent process-design co-development, even in spite of discontinuous
patterning, process, and device innovation is reiterated. The escalating design rule complexity resulting from increasing
layout sensitivities in physical and electrical yield and the resulting risk to profitable technology scaling is reviewed.
Shortcomings in traditional Design for Manufacturability (DfM) solutions are identified and contrasted to the highly
successful integrated design-technology co-optimization used for SRAM and other memory arrays. The feasibility of
extending memory-style design-technology co-optimization, based on a highly simplified layout environment, to logic
chips is demonstrated. Layout density benefits, modeled patterning and electrical yield improvements, as well as
substantially improved layout simplicity are quantified in a conventional versus template-based design comparison on a
65nm IBM PowerPC 405 microprocessor core. The adaptability of this highly regularized template-based design
solution to different yield concerns and design styles is shown in the extension of this work to 32nm with an increased
focus on interconnect redundancy. In closing, the work not covered in this paper, focused on the process side of the
integrated process-design co-optimization, is introduced.
Keywords: Design for Manufacturability (DfM), template-based design, design-technology co-optimization (DTCO),
predictably composable logic, pdBrix

1. INTRODUCTION
It is well known that profitability in the
microelectronics industry is driven by a two
year cycle in which transistor density doubles,
performance noticeably increases, power
consumption drops, and wafer manufacturing
cost remains largely constant. With
wavelength scaling reaching its limit at 193nm
in the 90nm node, patterning resolution
improvement has been achieved through a
series of discontinuous innovations rather than
predictable evolutionary enhancements like
wavelength reduction (Fig.1, top). Both
computational resolution enhancements, such
as off-axis illumination with sub-resolution
Figure 1 CMOS scaling is being challenged by radical innovations assist features and optimized model-based
in patterning, process, device, and interconnect technology, yet optical proximity correction, as well as
the pressure to stay on a 2 year node-to-node cycle remains. physical resolution enhancements, such as
ultra-high numerical aperture lithography
enabled through the use of water immersion, share the common trait of introducing more severe and more complex
layout sensitivities in physical and electrical yield detractors.
Beyond the aforementioned escalation in patterning complexity, the profitability of the microelectronics industry is
further challenged by the fact that dimensional scaling alone is no longer sufficient to achieve the electrical performance
targets of the next technology node. Additional device, interconnect, and process innovations, introduced with every new
node, further add to the unpredictability of layout sensitive detractors to yield ramp (Fig.1, bottom). Yet, even as process
optimization becomes more difficult and layout-specific, the two year node-to-node timetable requires concurrent
technology and design co-optimization. It is simply not possible to develop a new process solution and thoroughly
characterize all its layout sensitivities before starting the respective node’s design work. To have product designs
available when the process is scheduled to yield functional chips, design work has to proceed in parallel to the process
and device optimization work.

1.1. Design Rules

Traditional design rules had provided an elegantly simple solution that allowed process and design development to occur
concurrently rather than sequentially. Design rules, which were predominately derived from competitive scaling targets
but also involved a period of intense collaboration, negotiation, simulation, and optimization early in the technology
node, established process control targets that set goals for process development and provided rules to initiate layout
work. Even as the physical limitations of certain processes started to prevent equally aggressive scaling of all design
rules, rules could be published early in the technology node to allow designers to begin their work while process
engineers drove the process to the committed control limits. On rare occasion, the committed process control limits
could not be achieved, causing highly undesirable churn late in a technology node. In general, however, this approach
worked until non-monotonic layout sensitivities drove rapid escalation in design rule complexity (Fig. 2).
In the example in Fig. 2, ‘forbidden
pitches’ in the off-axis illumination
based patterning solution drive very
complex multi-feature width-dependent
spacing rules. Unfortunately, even with
all this complexity, these rules can not
provide absolute assurance that all
design-rule-clean layouts will yield or
perform adequately. Unanticipated
asymmetric width-space combinations
or two-dimensional constructs not
taken into account in the rule
optimization can easily cause yield loss
Figure 2 Illustration of the escalation in design rule complexity in response while passing the design rule checking.
to non-monotonic layout sensitivities. Shown is a typical example of On the other hand, some layout
minimum allowed metal spacing as a function of drawn and neighboring constructs that may be extremely
widths. What used to be a simple pass-fail limit has turned into a complex valuable to a particular design, may fail
multi-body problem. design rule checking but would be
adequately manufacturable. Rules-
based design simply has to make compromises between being too conservative and being too complex, which opens the
door for yield or performance challenges.

1.2. Traditional DfM

Two traditional Design-for-Manufacturability (DfM) solutions are in direct response to the challenge of design rules
becoming both too complex and too conservative; model-based optimization, and restrictive design rules.
Model-based layout optimization seeks to eliminate conservatism in the design by letting designers or design tools
interactively and iteratively detect and eliminate layout sensitivities by modeling the anticipated process response of a
particular layout construct. Illustrated in Fig.3a is the lithography response to a local wiring (M1) layout showing layout
sensitivities in terms of localized line and space narrowing. While in theory the concept of model-based layout
legalization may seem like a good idea, it is completely infeasible for all but a few niche applications: Of the large
number of process steps that each exhibit unique layout sensitivities, only lithography and chemical mechanical
polishing (CMP) have reasonably accurate process models. And even for these models, only the aerial image component
of lithography, not the chemical component of resist exposure, is based on predictive fist principle simulations, all other
effects are captured in semi-empirical models that require extensive experimental calibration. A partial set of empirical
models, calibrated to a rapidly changing process, clearly can not provide an accurate basis for layout legalization early in
a technology node. Predictive process modeling challenges aside, integrating model-based layout legalization into a
design flow would require substantial reengineering of the entire IP generation, synthesis, placement, and routing flow
and would severely err on the side of escalating design complexity in favor of reducing conservatism (1). The
noteworthy exception is lithography and CMP modeling based routing optimization late in the design flow (2,3).

Figure 3a Example of lithography variability bands Figure 3b Illustration of an RDR description of the
of a first-metal layout as they would be presented to width-dependent-spacing phenomenon captured in Fig 2.
a designer in interactive model-based DfM. Dots represent specific allowed line placements.

In contrast to model-based DfM, Restricted Design Rule (RDR) based DfM focuses on preserving, even enhancing,
design efficiency while ensuring yield and performance by eliminating unknown layout sensitivities. Best described as a
design approach based on ‘prescriptive’ design rules rather than the traditional ‘prohibitive’ design rules, RDR-based
design seeks to provide clarity in the design-process handoff by comprehensively defining all allowed feature
placements (4). As shown in Fig.3b, RDRs minimize the complexity of width-dependent spacing rules (Fig.2) by
defining discrete placement options for narrow lines, followed by a more traditional continuous design space with a
single conservative design rule for intermediate width lines, and complete elimination of all lines (with limited design
value) of extremely large dimensions. In a process environment where yield and performance no longer improve
monotonically above the minimum design rule, RDRs provide clear targets for aggressive process optimization. Key to
successful implementation of RDR-based DfM is a close collaboration between the design and process teams in the
initial definition of the RDRs to ensure process development for a limited set of design rules that is actually useful to the
designers (5). While currently the only feasible design-process co-optimization solution, RDR-based DfM still leaves
room for improvement in eliminating design conservatism while ensuring competitive performance and yield.
The common element of current DfM solutions is that, similar to the original design rules, they focus primarily on the
‘layout space’ after the actual design and before actual process optimization. Therefore, these approaches can only
provide sub-optimal solutions.

1.3. SRAM Design-Process Co-optimization

A look at how a logic chip’s memory arrays are developed for a new technology node
provides valuable insights into the next level of design-technology co-optimization. SRAM
designs have consistently pushed technology harder than logic, achieving very aggressive
transistor densities under very rigorous performance constraints. The aggressive scaling of
memory is achieved primarily through deep design-technology co-optimization: technology
options, including resolution enhancement techniques but extending to all aspects of
process, device, and interconnect development, are fine tuned to the unique requirements of
Figure 4 A SRAM the SRAM layout. As Fig.4 shows, this successful design-technology co-optimization is not
layout showing pattern enabled by forcing all designs onto simple one-dimensional gratings. Design rules are
complexity.
pushed to the absolute limit for each unique and often complex layout configuration to the
point where the design rule is replaced by the actual construct being scaled. Parameterized
variants of the SRAM cells are comprehensively qualified and characterized using dedicated test vehicles to ensure a
complete understanding of the yield implications of each specific layout construct. Multi-disciplinary engineering teams
interactively converge on an optimized layout and matching manufacturing process using rigorous simulation and
iterative experimentation. This comprehensive design-technology co-optimization is fundamentally enabled by a
simplified design space that allows competitive designs to be synthesized from a small set of predictably composable
logic constructs.

1.4. Maintaining Profitable Scaling

To assess whether an SRAM-like design-technology co-optimization is feasible and beneficial for logic designs in
advanced technology nodes, three key questions need to be answered:
- Can a competitive logic design be generated from a highly constrained set of logic constructs (templates)?
- Are these logic constructs really predictably composable?
- Does the constrained and predictable layout environment provide process optimization and characterization
benefits that result in improved yield and performance?
The work reported in this paper focuses primarily on the design’ability aspect of template-based design, though other
work, focused on demonstrating the broader design-process co-optimization benefits, is briefly introduced.

2. 65nm PowerPC 405s CORE DESIGN

2.1. Experimental Setup

To assess the feasibility of generating a competitive product design from a simplified set of layout constructs, IBM and
PDF collaborated in the redesign of a PowerPC 405s core in the 65nm technology node using PDF’s pdBrix design
methodology (6,7). Starting with the original netlist, the redesign was constrained to maintain the same:
- macro footprint and area (1111 mm x 1817.5 mm, including reused memory blocks)
- performance (3.5ns Clock Period, ~287 MHz)
- power (RVT / HVT transistor breakdown = 90% / 10%)

2.2. Fabric and Templates

To ensure predictable composability of the resulting logic constructs, the pdBrix design flow begins with the definition
of an ultra-regular cell image, referred to as the ‘fabric’, onto which all logic is mapped. Conceptually, the regularity of a
sample fabric is illustrated in Fig.5a, showing the extremely constrained design grid available for poly and metal shapes.
Fig.5b is a simple cartoon of a possible logic topology mapped onto this coarse design grid to illustrate that, while all
features have to fall onto the prescribed grid, not all grid lines have to be occupied by layout features. The combination
of this coarse layout grid and a handful of simple design rules yields a fabric that results in:
- limited diffusion corners
- fixed-pitch, unidirectional Poly (vertical)
- unidirectional metal (first metal, M1, horizontal; second metal, M2, vertical)
- relaxed pitch for M1 (17%) and M2 (25%)
- contacts and vias on grid
The relaxed metal pitch and the predictable nature of the gridded layout enabled the use of tighter rules for metal
extension past contacts and vias.

Figure 5a Figure 5b Simple Figure 5c Schematic Figure 5d Physical

Illustration of the logic mapped onto of a sample logic layout of the
Fabric showing uni- the Fabric, showing primitive, !(AB+C), Template defined in
directional fixed selective track use. representing a Fig. 5c.
pitch layout Template.
A specific example of the basic logic constructs used in this design is shown in Fig.s 5c and 5d. The schematic for a
!(AB+C) function is mapped onto the aforementioned fabric to yield the template layout shown in Fig. 5d.

2.3. Hotspot Reduction

The lithography benefits of ultra regular layouts are quite obvious. Not only does the fixed-pitch, single-orientation
layout simplify the frequency space that has to be imaged, it completely eliminates all complex two-dimensional
constructs that could lead to yield concerns. To confirm this theory, both the conventional and the pdBrix layouts were
scaled to 45nm dimensions and run through process-window lithography simulations. The scaling to 45nm was done to
provide a more challenging resolution environment typical of advanced technology nodes and to leverage more accurate
through-process lithography models available in 45nm. It provided the additional benefit of assessing the layout
migrate’ability differences in the two design styles. Fig.s 6a and 6b show the simulated lithographic process window
bands for the pdBrix and conventional layout, while Fig.6c shows examples of typical layout sensitivities discovered in
the conventional layout. The hotspot count in Table 1 shows the improved patterning robustness of the pdBrix design.
Litho Hotspots after scaling to 45nm

Poly Width M1 Width M1 Space

pdBrix stndrd pdBrix stndrd pdBrix stndrd

0 34 0 23 0 1634

Figure 6a Figure 6b Figure 6c Litho Table 1 count of lithography

pdBrix layout standard layout layout sensitivities: hotspots that introduce a potential
pinching and bridging yield risk for the standard and
Lithography simulations for: diffusion (red)
hotspots occurring pdBrix layout.
poly (green), 1st metal (blue), and contact
around 2-d constructs.
(purple) including dose, focus, and mask error

2.4. Variability Improvement

In addition to process limited yield (PLY) effects, advanced technology nodes are suffering from loss of circuit limited
yield (CLY); i.e. chips that don’t have any hard fails but don’t meet timing or performance requirements. A primary
factor in controlling CLY is the reduction of electrical variability. Fig.7 shows the improvement in transistor variability
achieved with the pdBrix layout. The lithography variability-bands for two layouts of the same logic function are shown
in Fig.7a. The variability-bands are comprised of individual contours that capture deviations in exposure dose, defocus,
and mask size. Fig.7b shows the electrically equivalent channel length extracted from these layouts based on the range
of process contours. Electrical channel length variability is reduced from 4.1nm to 3.5nm, which is equivalent to as
much as 10-15mV reduction in threshold voltage variation. This translates into a very significant improvement in CLY.

Figure 7a Diffusion, poly, and contact Figure 7b Range of electrical channel length extracted for each
lithography variability-bands for pdBrix layout transistor in the two layouts of Fig 7a. Increased electrical
(top) and conventional layout (bottom) variability as a response to lithography variation due to dose,
focus, and mask size is seen in the standard layout.
2.5. Area Improvement
While the yield and performance benefits of regularized layouts may be well accepted, the biggest barrier to broader
implementation of regularized layout styles is the perceived impact on layout density. The pdBrix design of IBM’s
PowerPC microprocessor showed that it was possible to contain the extremely regularized layout in the same footprint as
the original layout. Further, the total area occupied by sequential logic was identical in the two layout styles and the area
occupied by combinatorial logic actually decreased by 25%. Further analysis of the specific contributors to this 25% area
reduction indicated that:
- 25% of the reduction was achieved through Template and Fabric co-optimization; i.e. elimination of layout
conservatism by implementing construct-specific design rules.
- 5% of the reduction was achieved through the use of design specific complex gates, referred to as Brix,
synthesized from the primitive logic functions rendered in Templates.
- 70% of the reduction was attributed to optimal construction of application-specific logic cell functions; i.e.
eliminating conservatism in layout variants like power levels by synthesizing an application specific library
from a technology node specific set of Templates rather than using generic set of standard cells designed to
cover all possible power/performance needs.

2.6. Design Simplicity Improvement

Reduction of lithography hotspots and improvement in
variability were expected benefits of highly regularized
layouts. Very effective use of the physical and electrical
predictability of the highly regularized layout style to
eliminate layout density penalties was necessary to make
this design approach possible. But the core value
proposition hinges on demonstrating the feasibility of
generating a competitive design from a substantially
simplified set of layout primitives. Fig.8 compares the
number of unique logic cells used in the standard
microprocessor core design to the number of cells used in
the pdBrix version of that design. The cell count is
reported as a function of the number of logical inputs to
the operation, showing that the pdBrix methodology
shifted the design towards more complex and compact
logic functions. The small number of logic cells common
Figure 8 Substantial simplification of the design space to both design styles further emphasizes that the
as measured by the number of unique logic cells used in reduction in cell count is fundamentally enabled by the
the standard versus the pdBrix design.
optimal use of application specific logic cells, not merely
a constrained physical synthesis.

3. Fabric Adaptability
The 65nm PowerPC design exercise:
a) confirmed the inherent patterning benefits of regularized layouts;
b) demonstrated that an optimally integrated design flow can achieve competitive layout density with a highly
regularized layout style; and
c) showed that it is possible to generate a competitive design from a substantially simplified set of logic
elements.
Based on these positive results, it was decided to continue this work in the 32nm node, driving the project closer to the
leading edge of the technology node to be better aligned with new product design starts. In porting the 65nm fabric to
32nm, three opportunities to align the Fabric more with product specific internal design objectives were identified:
 - Running first metal perpendicular to the poly gates and forcing it to be
completely uni-directional eliminates all possibility of multiple diffusion contacts,
as shown in Fig.9A. It is well known that redundant diffusion contacts are
preferred in some designs for device performance and yield benefits. Further, the
strictly unidirectional metal adds additional vias for all ‘wrong’ way connections,
which is not ideal for processes with via yield challenges.
Figure 9 Illustration of 3 - Forcing all metal to be equal width (Fig.9B) eliminates the possibility of using
specific layout considerations wide wires for power distribution, preferred in some designs to improve reliability
in the original pattern-count on these high current carrying constructs.
optimized Fabric: contact
redundancy, narrow power - Linking the contacted device pitch (i.e. minimum poly pitch) to the minimum
wires, tight tip-to-tip space. contacted tip-to-tip spacing of first metal (Fig.9C) creates a scaling problem.
While pitch scaling is driven to 30% reduction per node, tip-to-tip spacing has
been scaling at roughly 20% per node, forcing a decoupling of these two constructs in the overall density scaling.
In a demonstration of the adaptability of the pdBrix design approach, a new fabric with different optimization priorities
was defined, as shown in Fig.10. The new fabric continues to enforce: limited diffusion corners, fixed-pitch
unidirectional poly (vertical), preferred orientation metal (M1 now vertical, M2 now horizontal), contacts and vias on
grid, and relaxed pitch for M1 (25%) and M2 (5%), but also accommodates limited wrong-way metal (on-grid) and
wider power rails. The improvement in via and contact redundancy for different layout options is shown in Fig 11. The
initial improvement is achieved by exploiting opportunities to insert redundant contacts and vias into the original image
(labeled ‘plus redundancy’ in Fig.11), followed by allowing limited use of wrong-way metal in the original image, and
finally by switching to a new cell image. Fig.11 compares the ‘number of non-redundant connections’ for all four cases.
Reporting interconnect improvement in this fashion highlights the fact that both ‘eliminating connections’ and ‘adding
redundancy to connections’ fulfills the stated design intent.

Figure 10 Fabric chosen to optimize Figure 11 Reduction of non redundant connections, relative to the
redundancy, reduce necessary connections, original PowerPC design, based on different fabric trade-offs.
and allow wide power wire.

Figure 12 A nand2 rendered in the ‘low Figure 13 Long range parametric layout sensitivities on diffusion and
pattern count’ (left) and ‘high redundancy’ poly are minimized through macro regularity (left), local yield hotspots
(right) Fabric on first metal are prevented through control of boundary conditions.
The differences in the two fabric styles, one optimized to reduce overall pattern count, the other optimized for
redundancy and wide power rails, can be seen in Fig.12. Shown are two templates representing the same logic gate (a
nand2) mapped onto the two fabrics. Defining the fabric with selective wrong-way metal does not affect the layout
density or logic simplicity as demonstrated above, but stresses the need to actively manage predictable composability of
the logic elements. To ensure predictable circuit performance regardless of placement, layout sensitivities in the active
device parameters have to be minimized. In addition to local proximity effects like poly line-width variation and
diffusion corner rounding, device performance is affected by long range layout sensitivities in processes such as etch,
stress, or rapid thermal anneal. These long range effects are best controlled through macro-regularity; i.e., global pattern
uniformity on diffusion and poly shapes as shown in Fig.13(left). The first metal level, being used for local interconnect
only, is not a major contributor to parametric variability, making local patterning hotspots the primary yield concern.
The small number of logic constructs allows the safe use of optimally complex layout configurations as long as proper
boundary conditions are enforced to ensure that no new complex layout configurations can be formed at cell boundaries.
As illustrated in the right layout of Fig.13, in this fabric the more complicated metal patterns (i.e. staircases or dense T-
shaped line-ends) are constrained to the center of the cell, as are ‘belt buckle’ constructs that require tight dimensional
control since they can be used for connectivity to other layers. Layout constructs within the boundary region of the cell
(i.e. outside of blue outlines in Fig.13) are simpler and more conservative (e.g. line-end connectivity in boundary region
belt buckles is not allowed). Predictable composability is thereby maintained for two layout sensitivities of different
length scales through two different mechanisms: macro regularity to address long range effects and boundary conditions
to address local effects.

4. Integrated Design-Process Co-Optimization

While the work reported here focused on the design and layout aspects of the pdBrix solution, these design studies are
complemented by integrated testsite runs under way in both 32nm SOI and bulk technologies. Experiments have been
designed to demonstrate:
- variability and design-margin reduction afforded by the predictable and regularized layout environment
- yield and manufacturability improvement facilitated by the simplified design environment
- predictable template composability achieved through layout regularity and control of boundary conditions
- performance and power benefits from the use of complex logic gates
- placement agnostic delay achieved through macroscopic layout regularity
- model-to-hardware correlation improvements through pre-characterized templates
In addition, the power/performance/area benefits, route’ability, and design flow compatibility of the pdBrix solution is
being confirmed in another design experiment targeting high performance macros from IBM’s Cell processor design.

5. CONCLUSION
It was shown that profitable CMOS scaling under the time pressures of the established two year node-to-node cycle in an
environment of continuous disruptive technology innovation requires deep design-technology co-optimization. To
overcome the shortcomings of DfM solutions that operate primarily in the physical layout space and do not address
fundamental design optimization or design-aware process optimization, the pdBrix design methodology was evaluated.
The goal of this collaborative work is to establish a comprehensive design to silicon solution that facilitates rigorous
design-technology co-optimization.

The key requirements for an optimal design to silicon solution have been establishes as:
- late binding of logic to physical layout to preserve design-creativity and -performance
- optimally simplified layout to allow construct-driven process development
- targeted characterization and qualification test vehicles to drive yield ramp

The work reported here demonstrated feasibility of design with a simplified set of logic, preserving layout density and
improving patterning yield and variability while achieving the stated power/performance targets with significantly fewer
layout constructs.
 ACKNOWLEDGMENTS

The joint IBM and PDF team of authors thanks their many colleagues that were directly and indirectly involved in this
work and acknowledges the specific contributions of Henning Haffner (Infineon) for his work on 32nm design rule
optimization and Dureseti Chidambarrao (IBM) for his work on contour-based extraction.

REFERENCES
1
DFM lessons learned from altPSM design, L. Liebmann, Z. Baum, I. Graur, D. Samuels, Proc. SPIE 6925, 69250C (2008)
2
Convergent automated chip-level lithography checking and fixing at 45 nm, Valerio Perez, et al. , Proc. SPIE 7275, (2009)
3
Hotspot detection and design recommendation using silicon-calibrated CMP model, Colin Hui, et.al., Proc. SPIE 7275, (2009)
4
Layout impact of resolution enhancement techniques: impediment or opportunity?, L. Liebmann, ISPD’03, 2003, Monterey
5
Intel design for manufacturing and evolution of design rules, Clair Webb, Proc. SPIE 6925, 692503 (2008)
6
Regular Fabrics for Nano-Scaled CMOS Technologies, L. Pileggi and A.J. Strojwas, ISSCC, (2006).
7
Maximization of Layout Printability/Manufacturability by Extreme Layout Regularity, Tejas Jhaveri, et al.,, J.
Micro/Nanolithography, MEMS, and MOEMS, Vol 6 (03), 2007.