Advertiser and

Buyer's Guide
Buyers Guide

Chip Shots blog

Greatest Hits of 2005
Greatest Hits of 2005

Featured Series
Featured Series

Web Sightings

Media Kit

Comments? Suggestions? Send us your feedback.


Can't we all just get along? Designers and manufacturers grapple with DFM


Although much progress has been made over the past few years to bridge the traditional gaps between the chip design and manufacturing communities, some of those involved can't even settle on what the ubiquitous initialism DFM means: Is it design for manufacturing (the most common definition) or design for manufacturability (perhaps the more accurate one)?

Sure, the designers say, our design can be manufactured; all of our models and previous experience tell us so. Right, says the fab team, but will it yield well, and at what cost and level of process complexity? With the advent of 90-nm device technologies and especially the looming 65- and 45-nm nodes, the two sparring groups must start to speak the same language, communicate better, and learn from each other, because the stakes are high. This issue's roster of Hot Button experts explores some of the critical issues surrounding DFM.

BILL VOLK (senior director of marketing, DesignScan, KLA-Tencor): The driver for DFM methodologies is the dominance of limited yield loss from systematic mechanisms. This trend leads to lower initial yields in manufacturing and lower mature yields, which tend to worsen with each process generation, despite ongoing advances in metrology and yield management. Systematic, pattern-related yield loss now dominates random yield loss, and with further investigation, we find it's all feature-limited yield, which relates back to design. That's such a big piece of yield loss, if we don't knock that out of the equation, we can't solve the performance issues going back to place and route.

Resolution enhancement technology (RET) teams need to verify that each design will yield before making reticles or silicon. Engineers can't continue to troubleshoot design issues in manufacturing, because the cost to cycle time is too high. We also hear interest in the electrical critical path being fed downstream, but we're pretty far from that. First, we must take care of the functional yield, and then solve issues causing the chip to miss binning targets. Once we correct the features that cause systematic yield loss and degrade the performance of the chip, maybe we can get into optimization of critical paths.

We must get a handle on manufacturing process variability in order to create a calibrated model of the variability so that information can be translated to the design side. In the case of lithography, this "lithography-aware design verification" involves inspection of the design data with advanced models to determine how the device pattern is going to perform throughout the lithography process window. To implement this, it takes advanced lithography knowledge with CD-SEM expertise and resources to perform lithography calibration and then tie accurate simulation models to the design for lithography process window characterization.

Most electronic design automation (EDA) tools for mask synthesis require that RET design teams do the calibration, relying on information from manufacturing. Our observation has been that the handoff of information coming from manufacturing is highly questionable unless you have deep insight into lithography. Otherwise, there are a lot of false starts on the calibration of models for process simulations.

"Design-aware process control" will be the next focus of DFM, such as identification of hot spots, which comes out of the optical proximity correction (OPC) tools. Hot spots are typically areas of features that could not be decorated because of restrictions on the OPC model due to the physical layout. These locations are fed downstream as thousands of areas, and if you have insight into which features are going to fail, not based on rules but on actual simulation, you can optimize the inspection and metrology sampling to areas that will be affected first by lithography process variations.

In summary, DFM cannot be composed of a single-point solution. The linking of manufacturing process with metrology and inspection results to upstream design is the key to success. The critical first step is creating a mechanism for lithography-aware design verification by providing accurate and painless lithography models for design. For design features that are still limiting a manufacturing process, the implementation of design-aware process control will provide an economical approach to monitoring manufacturing process variability and yield excursions.

JOE DRAINA (associate director, International Sematech Manufacturing Initiative): The integration of design with manufacturing has been a staple in many industries, but it's been something of a latecomer to the semiconductor world. Until fairly recently, design generally drove processing, but many new factors—such as increasing variability, mask costs, exploding data availability, and litho hardware limitations—are posing significant challenges for IC designers. One response had been the emergence of DFM, which aims to address productivity early in the design cycle, so that manufacturing issues and concerns are integrated with design to obtain more-producible products.


It is important for R&D and
design to consider a design's manufacturing affordability.
—Joe Draina

My experience with DFM has been in the area of process equipment manufacturing readiness and robustness. My recent experience as senior manager of equipment engineering at IBM's 300-mm facility in East Fishkill, NY (where I was responsible for fab equipment ramp-up methodology, implementation, and process equipment productivity) has demonstrated how important it is for R&D and design to strongly consider a process design's manufacturing affordability. Here are some examples:

Wafer processing time and its effect on equipment throughput. Consideration should be given to etch or deposition rates, implant time, furnace residence time, CMP removal rates, added recipe overhead, additional purging steps for cooldown or particle control, extra rinse steps, and the like. These factors can add cost and complexity. Lower-than-anticipated throughputs cause capital costs to increase because of the need for more equipment and fab space. Accordingly, cost of ownership (COO) must be considered before commitments are made to manufacturing implementation.

Chemical costs and increased equipment complexity. The introduction of a significant toxic, pyrophoric, or flammable gas to a process recipe will cause total capital equipment costs to rise at least 10%, because of the added expense of double-wall stainless pipe delivery systems, additional scrubbers, safety interlocks, and alarm systems. The increase in equipment com-plexity means that more maintenance skills and headcount will be required to service the system. Expensive gases and chemicals, such as NF3, silane, helium, some slurries, and photoresists, should be used minimally or replaced with cheaper alternatives.

Chamber recovery to qualification specs after being opened to atmosphere or wet cleaned. Attention to this area is needed for any new vacuum process implementation. In particular, RTP gate hot processes and some insulator processes can be very sensitive to moisture that can getter process reactants. Defining these process recipes in conjunction with chamber-seasoning parameters can be critical since a poor relationship can result in seasoning times taking one to several days. Many etch and deposition processes require seasoning after the chamber is exposed to atmosphere or wet clean, so seasoning should also be included in the process design here. The periodicity of chamber opening or reactor wet cleaning must be considered. Pursuing changes in this area while running in manufacturing can be especially troublesome, since equipment can be sidelined for extended periods of time.

Consumable parts costs and frequency of exchange or wearout. Any new process may drive different consumables or higher wearout rates for elements such as electrostatic chucks, "wetted" chamber parts, scrubber systems, and liquid-delivery or gas-flow systems. Sometimes a seemingly minor adjustment to a reactor wall temperature or gas flow causes a higher deposit rate on chamber sidewalls, resulting in particle performance degradation, earlier O-ring degradation, electrostatic chuck arcing, greater frequency of wet cleaning, and longer seasoning times. COO must be understood relative to any parts change or increase in usage, as well as equipment downtime associated with frequency of exchange.

The above are examples of areas to consider prior to making commitments for process design implementations in a fab. The challenge for designers is to adequately investigate these areas before making manufacturing commitments, in order to minimize or eliminate activity in these areas while running a manufacturing fab.

JEAN-MARIE BRUNET (LFD market development manager, design to silicon division, Mentor Graphics): The ability to accurately print the image intended by the designer is one of the greatest challenges at the 65-nm node. This is especially true given the extensive use of RET at this node. But modifications to a photomask are no longer sufficient to ensure image fidelity. That's why foundries, designers, and EDA tool providers are turning to DFM methodologies that promise improved flows and solutions for managing yield.

Traditionally, design-rule checking (DRC) acted as one of the key communication vehicles between manufacturing and design, informing the designer of limits imposed on them by manufacturing. Most constraints represent true process limitations, which, if not followed, produce nonfunctioning silicon or considerably low yields.

Foundries, designers, and EDA tool providers are turning to DFM methodologies.
—Jean-Marie Brunet

Recently, foundries have also started delivering "DFM recommended rules," which indicate where a design becomes easier to manufacture by adhering to the DFM rule, rather than the minimum-spacing DRC rule. The quest to comply with DFM recommended rules opens the door to many questions: for example, what tools help the designer determine if DRC rules or DFM rules have a more positive impact on yield? One way is to gather layout statistics on the feature in question. For instance, statistical modeling and critical-area analysis highlight not only how often an issue occurs, such as antennas and vias, but also in which combination and level of severity the issue occurs.

While DFM-recommended rules go a long way to mitigate some yield issues, at 65 nm there is an even greater issue: Designers also need to assess how physical layout, especially at the cell library level, can be done so that feature fidelity is preserved across the manufacturing process window, not just at nominal dose and focus.

New developments in lithography-friendly design (LFD) methodology are making the goal of modeling process variability to improve layout robustness a reality. With an LFD "process kit" that encompasses RET recipes, process models, and parameterizable rules, designers can run simulations to see how a layout will print across the lithographic process window. Simulation results can include recommendations about areas where modifications to the layout would most likely improve yield, with modifications made in the native layout environment. While the kit contains all the data pertaining to pattern transfer, what the designer sees and works in is very much like a DRC environment. As designers become used to working in LFD mode, they will learn what design elements respond favorably to manufacturing processes, and, in time, naturally achieve a "LFD clean" design.

ARTUR BALASINSKI (engineering manager, process technology R&D, Cypress Semiconductor): If I were to look into a crystal ball, the recent Photomask Japan conference provided a perspective on what solutions may be in store for design for yield (DFY) challenges. For lithography, there appear to be two trends: one for the immediate term and one for the ultimate term.

The immediate trend is sweaty. It is not necessarily cheap but definitely less expensive than what's coming next. It requires a lot of effort from design, computer-aided design (CAD), and technology to benefit from the low-k1 resolution factor and stepper numerical aperture (NA) stretched to the limits. Model-based 193-nm OPC for 45-nm random logic has every chance to skyrocket mask complexity to multiple gigabytes and data processing time to many days. Lithography engineers will be busy optimizing exposure models and harassing design engineers about layout best practices to enable half- or quarter-pitch printability, slowing down data delivery, but making EDA engineers work twice as hard on new flavors of scatter bars. The reduction of unnecessary layout notches, definition of forbidden pitches and orientations, and addition of enclosures for line-end vias spiral up within the existing methodology, but to help out the cost, the expertise would eventually transfer to Asia.

However, this era should not last for long. The IC industry leaders are tired of paying millions for upgrades in stepper NA that are only good for one generation. They want to leverage their resources toward extreme UV, direct write, or imprint. They are not ecstatic about these choices: the throughput of new tools may be half that of the good old 193-nm stepper, but their risk is low. Moore's Law is alive and kicking, if one looks at the International Technology Roadmap for Semiconductors. They want the new lithography ready in a few years. This is a fantastic challenge for the stepper vendors.

My crystal ball tells me it is going to be EUV, but it does not really matter. Any solution should trash the existing OPCs. The IC industry would happily trade the software rigmarole for the hardware fix. The new litho tool would be the Noah's ark, which allows the IC companies that can afford the ticket to sail over the sea of OPC solutions. Masks would become less complex, if not less expensive. But many of the great EDA engineers may need to find new jobs. Perhaps they can help the designers who are furious about litho-oriented DFM focus with its other aspects (pattern density, wire spreading, or doubling vias for reliability) that are closer to their core expertise, such as timing analysis or leakage reduction. The resolution of these issues does not represent a fundamental technical challenge but does require some diligence, programming skills, and keeping an eye on one's return on investment.

What about smaller companies? They should expand existing OPC expertise and improve its cost-effectiveness. When they also acquire EUV, they should be able to offer a competitive solution to the multiple ASIC and R&D centers in search of innovation. If at least a few chosen ones have leading-edge litho tools, this will allow others to avoid or delay investing big capital dollars.

DIPU PRAMANIK (director of DFM solutions, Synopsys) and TERRY MA (director, product marketing for DFM solutions, Synopsys): A key requirement for DFM is a mechanism to incorporate process information that affects the functionality and yield of designs into the flow so that the target yield at the start of mass production can be attained quickly. With shrinking process tolerances, however, it is becoming more difficult to achieve the yield goal and, at the same time, meet the time-to-market constraints.

TCAD complements silicon data with accurate process and device simulations.
—Dipu Pramanik

At 90-nm and subsequent technology nodes, parametric yield loss becomes a significant part of overall yield loss because of the sensitivity of design parameters (frequency, power, etc.) to process variations (gate oxide thickness, gate critical dimension, halo implant dose, etc.). To solve this problem, a bidirectional link between design and manufacturing is needed. Design must account for manufacturing variations in a way that does not sacrifice the performance advantages provided by the new technology, while manufacturing must incorporate design constraints into the overall process control structure.

Technology computer-aided design (TCAD) provides the bidirectional link between design and manufacturing because it offers a predictive framework for correlating design parameter sensitivity with process variability. TCAD complements silicon data with accurate process and device simulations based on calibrated models. Extensive TCAD simulations with calibrated models allow the user to capture the relationships between individual process parameters and key device or even design characteristics and cast them into efficient and fast mathematical models called process compact models (PCMs). PCMs provide a robust way of transferring process and device information into both design and manufacturing.

In the design flow, PCMs can be incorporated into timing analysis and physical verification tools to determine the sensitivity of key circuit elements to process variations. In manufacturing, PCM enables advanced process control methods that can reduce the overall performance variation of a design caused by the random statistical variations of individual process steps.

The methodology involves feeding CD, gate oxide, and other in-line metrology measurements into the PCMs to determine the value of subsequent process parameters. For example, if the final gate CD is at the lower end of the specifications, with PCMs the user can calculate the change in halo implant dose needed to keep the off-state leakage current within limits while keeping the on currents in specification. It is also possible to distribute the change among several different process parameters, such as dose, implant angle, and anneal temperature. The net result is an improvement in parametric product yield for specific designs without the manufacturing engineer having to understand the design details.

MicroHome | Search | Current Issue | MicroArchives
Buyers Guide | Media Kit

Questions/comments about MICRO Magazine? E-mail us at

© 2007 Tom Cheyney
All rights reserved.