The standard cell library industry is in a shakeout. Although many libraries are given away "free" with processes, they are still expensive and labor-intensive to build. This makes it difficult for anyone in the library business to provide the value that customers expect and at the cost they expect to pay.

At the Design Technology Center for Agilent Technologies' Imaging Electronics Division, we have taken measures to drastically decrease the amount of manual labor involved to create libraries, reducing the cost to competitive levels and opening up new possibilities for the role of standard cell libraries. Our team has adopted a new, fully automated methodology for creating libraries that will help us achieve this goal.

The question that often arises when discussing whether to build libraries is, "Why not simply purchase standard, off-the-shelf libraries?" As processes become more similar, the value in creating a custom library isn't immediately obvious. If the library can be made with negligible time and effort compared to the cost of buying one, however, the value is clear. An analogy to buying bread applies here. Why would someone take the time and effort to make their own bread when they could easily buy it off the shelf at a supermarket? But if the bread could be made at home without expending more time, effort or cost than it would take to buy off-the-shelf bread, the advantages become more significant.

For the same cost and effort as off-the-shelf products, we not only create custom libraries comparable to off-the-shelf offerings, but we can also create new, custom libraries that weren't otherwise available. Custom ability at commodity pricing-that was our goal at Agilent when we started investigating automated library manufacturing. One year later, we have an automated library system that provides everything we need.

Automated versus manual library creation

The advantage to custom library creation is clear, but there is still a question of how to create those libraries. To achieve the custom capability while still keeping the cost, effort, and time comparable with the ease of buying off-the-shelf libraries, we need to explore the capability of automated library creation. With the time-to-market pressures and tight design turnaround times needed to meet the market's demands, manual library creation is no longer feasible.

A manual library creation methodology is too involved to meet current cost and schedule requirements. The manual method takes a large team of engineers devoted to the process over long periods of time. It's expensive in terms of both cost and resources and, most importantly, it simply takes too long to fit into today's accelerated design schedules. Despite the advantage of high quality libraries that the manual creation method provides, designers can't consider it a productive method for library creation in the future.

Assuming that we want to pursue custom library creation, we must examine the automation of this process, since automated library creation has both advantages and disadvantages. At first glance, it seems that the obvious advantage of automation over traditional manual creation methods is greater efficiency, which, in turn, results in lower cost, and libraries that are available sooner (see Figure 1). It seems the obvious disadvantages are resulting cells that are less efficient than handcrafted ones. This would be the case if the efficiency difference were small, say on the order of two times more efficient for automated library creation. However, if the efficiency differences are quite large-closer to two complete orders of magnitude more efficient-some remarkable advantages emerge.

Figure 1 - Automating library creation

Classic-SC, an automated layout tool, creates handcrafted layout from SPICE netlists or from existing layouts.

The most substantial benefit is that more than one type of library can be created per process. The effect is that we can balance a trade-off made to sacrifice individual cell efficiency by targeting entire libraries to more efficiently cover the customer design space. For example, if a developer were limited to one library, how would he or she determine the best library to use? Would the decision be based on the characteristics of the individual cells? Perhaps, but the developer would be more interested in the capability of the entire library to meet design needs.

If you needed to design a 400 MHz CPU core, you couldn't use a library designed for 300 MHz operation, no matter how efficiently each cell was created. Similarly, if you were looking for extremely low-power operation at 50 MHz, you would not choose a 300 MHz library either. Perhaps you would choose an entirely different process as well. The automation of library creation enables the development of more library solutions to meet the diverse needs of individual projects.

At this point, you might think that with the availability of multiple libraries from multiple vendors, you could still gain this advantage with off-the-shelf libraries. While it's true that many different types of libraries are available for many different types of processes, these products are all extreme in their diversity. Any design center relying on a particular set of tools, test methodologies, and documentation styles will find that all of those elements differ greatly for each different library.

If, for example, a designer normally compares cell performance with worst-case numbers, it will be difficult for them to use documentation with only typical cases. Design centers don't want to learn completely new creation flows for each design. After consideration, you can see that off-the-shelf libraries are still not the optimum solution. It would take longer to make all the off-the-shelf pieces fit each time and there would still be no way to create the occasionally required custom cell.

One of the most exciting prospects for the automated cell manufacturing system is the ability to create fully custom cells as quickly and easily as any other cell and have it work identically to every other library cell in terms of tools and methodology. Once this is achieved, designers gain all the advantages of the manual creation method, without the drawbacks of time and cost.

Assembly line library automation

At Agilent, our first step in creating the automated library system was to study the details of the current manual library creation process. Our intent was to retain the high quality enabled by the traditional methods for creating libraries while making a system that could be automated. Our plan was to take all the steps that had gone into creating libraries, eliminate redundancies, and rearrange the steps into a linear flow from specification to manufacturing.

We then broke up these steps into several different modules-separate pieces of code that operate on cells individually and independently. We put the whole system together like an assembly line for cars. Each cell is processed through each module until the cell has passed all the steps and is ready for shipment. The main difference between the automated manufacturing engine and a typical manufacturing assembly line is that each cell has its own individual "virtual assembly line." This method allows all cells to be processed at the same time, in parallel.

The advantage of processing all cells concurrently is that the entire library could theoretically be completed in the time it takes to complete the most complicated cell. In reality, this time is limited by the amount of licenses that can be acquired for certain tools, the amount of processing power available, and other factors which may not be as obvious.

The first of these other factors is the limit to how fast engineers can design complicated cells. While scaling may be a quick and useful technique for simple combinatorial cells, other cells that contain analog-like behavior, such as feedback mechanisms, need to be carefully tailored to each process. At Agilent, we experienced that situation in the libraries we created, with combinatorial cells designed and manufactured at the rate of hundreds per day while the much more complicated scan flip flops took two or three days for each cell.

The second factor is that we can't create any cells before the entire manufacturing system is ready. It's difficult, for example, when we try to develop both the module that creates layout and the module that checks layout correctness (design rule check, DRC) at the same time. It becomes a chicken-and-egg problem, with the final debug and test of each module waiting for the final debug and test of the other modules.

We found that close communication and coordination between all module developers was required in order to realize the productivity gains from the automated generation system. The main reason this becomes such an issue isn't because of the automation system itself, but the expectation of tighter schedules that the automation system produces. When the schedule allows for nine months to create a library, those issues can be worked out with relative ease, but when the goal is to build a full library in two months, managing the coordination of all parts becomes more critical.

Putting the process in motion

The automated library manufacturing system consists of a number of modules and an "umbrella" program that controls the operation of the modules on the cells.

The umbrella program constantly monitors the status of all cells. When a cell is ready for processing, the umbrella program "sweeps" it into manufacturing and starts executing the modules on it one at a time.

As one module completes its operations on a cell, the umbrella program is alerted and then starts the next module on the cell. The umbrella program is capable of handling operations for each cell individually in real time on all cells in the library, unaided by any human effort.

The following outline describes the major steps we took for the processing of individual cells:

Schematic verification-Schematics are used as source specifications for each cell in the library. Each schematic is verified for completeness and functional correctness with our proprietary switch-level simulator, AWSIM.

Layout creation-Layout is created automatically for each cell through Cadabra Design Automation's (San Jose, CA) third-party tool, Classic-SC. This tool generates artwork from inputs of library and cell netlist specifications. The following section explains why we chose the Cadabra tool.

Layout Verification-Layout is verified for correctness with currently popular DRC tools (our proprietary Trantor, or other commercially available ones). Layout is also verified against the source schematics (LVS) to make sure the intended circuit was created accurately (with our proprietary Hverify, or other commercially available tools). Library consistency checks (LCC) are also performed to make sure that each cell conforms to a standard for the library and that it can be successfully used in conjunction with the other cells in the library.

Characterization-Cell performance and operation is verified and measured with SPICE simulation (using our own proprietary HPSPICE) and this information is used for the creation of models that are released for each cell.

Bringing in a third party

As stated above, we used the Classic-SC tool from EDA vendor Cadabra. It's worth mentioning this tool specifically, since a fully automated layout tool is the most critical element for making library automation work. Cadabra offers an automated flow called the Library Factory, a method for layout creation that enables rapid development of custom cells and libraries, and ensures that these cells meet sub-wavelength process technology requirements. The tool was the only third party tool we needed to use in our library creation system, plugging into our own custom Library Factory that we assembled with tools developed in-house at Agilent.

I should also note, however, that it was not strictly necessary to use all our own tools. For example, we have also used third-party DRC/LVS tools in addition to our proprietary tools in our automated system. Companies that don't have the resources of an in-house CAD group could also be successful in automating the creation of their own libraries.

Figure 2 - Automated Transistor Layout (ATL)

An advantage of the tool is the automatic creation of LVS- and DRC-correct layouts for standard cells.

The tool is based on Cadabra's Automated Transistor Layout (ATL) technology (see Figure 2). The advantage of this tool is the automatic creation of LVS and design rule correct layouts for standard cells. Starting with a SPICE level netlist or an existing GDSII layout, and using physical specifications that describe the common attributes of the library's cells, the tool creates cell layouts with a level of quality that was previously only possible with the painstakingly detailed work afforded by manual creation.

Features we liked about the tool also include the placement capability that groups, arranges, and stacks transistors to maximize diffusion sharing and routability; the dedicated, customizable router with support of 45-degree routing; the intelligent two-dimensional compactor, providing full support of all 45-degree shapes; and support for advanced design rules.

Another advantage of the tool is that it can handle incorporation of optical proximity correction (OPC) and phase shifting techniques, which make the layouts sufficiently complex so that they can no longer be created manually. This is an important issue today as process technology advances and transistor channel lengths shrink below 0.18 micron. Subwavelength chip designs generate inconsistencies between the layout, mask, and actual silicon, which can ultimately impact the design quality. OPC and phase shifting techniques can be used to resolve these inconsistencies, but the resulting library complexity makes an automated tool even more essential.

Achieving successful results We recently used the automated library manufacturing system at Agilent to create a TSMC 0.18-micron standard cell library. It took only two months to complete this library of over 300 cells, including all cell design and tool customization.

Figure 3 - Ring of a tree

The project started January 4, 2000.

The cell generation gap displays data collected during the creation of this library (see Figure 3). Almost like rings on a tree, this graph indicates the phases of our library development. Initially, we used the first 50 cells to create the TSMC 0.18-micron process customizations and to test the system. When the customizations were sufficient to use the full automation system, we completed the bulk of the combinatorial cells in about three days. After that, it appears the automated system slows down considerably, but this is more an indication of our inability to design complex cells fast enough to keep the system working full time. More tool customization was occurring during this time and required both a full recharacterization and recreation of LEF models at different times. Neither of these efforts affected our cell creation efforts since they were handled automatically by the system.

In conclusion, the move toward automation of library creation is a necessary migration induced by the growth and complexity of chip designs. The ability to create quality custom libraries rapidly and inexpensively will give chip designers a major competitive advantage over the next several years, as evidenced by the dramatic speed with which we developed our custom libraries. Using in-house tools or third-party tools that are available today can help designers develop an automated library generation method to deliver custom ability at commodity prices and schedules for all chip designs.

David Pietromonaco is in the imaging electronics division of the Design Technology Center for Agilent Technologies (Palo Alto, CA).