- 1.1 Why Design Integrated Circuits?
- 1.2 Integrated Circuit Manufacturing
- 1.3 CMOS Technology
- 1.4 Integrated Circuit Design Techniques
- 1.5 IP-Based Design
- 1.6 A Look into the Future
- 1.7 Summary
- 1.8 References
- 1.9 Problems
This chapter is from the book
This chapter is from the book
This chapter is from the book
1.4 Integrated Circuit Design Techniques
To make use of the flood of transistors given to us by Moore’s Law, we must design large, complex chips quickly. The obstacle to making large chips work correctly is complexity—many interesting ideas for chips have died in the swamp of details that must be made correct before the chip actually works. Integrated circuit design is hard because designers must juggle several different problems:
- Multiple levels of abstraction. IC design requires refining an idea through many levels of detail. Starting from a specification of what the chip must do, the designer must create an architecture which performs the required function, expand the architecture into a logic design, and further expand the logic design into a layout like the one in Figure 1-2. As you will learn by the end of this book, the specification-to-layout design process is a lot of work.
- Multiple and conflicting costs. In addition to drawing a design through many levels of detail, the designer must also take into account costs—not dollar costs, but criteria by which the quality of the design is judged. One critical cost is the speed at which the chip runs. Two architectures that execute the same function (multiplication, for example) may run at very different speeds. We will see that chip area is another critical design cost: the cost of manufacturing a chip is exponentially related to its area, and chips much larger than 1 cm2 cannot be manufactured at all. Furthermore, if multiple cost criteria—such as area and speed requirements—must be satisfied, many design decisions will improve one cost metric at the expense of the other. Design is dominated by the process of balancing conflicting constraints.
- Short design time. In an ideal world, a designer would have time to contemplate the effect of a design decision. We do not, however, live in an ideal world. Chips which appear too late may make little or no money because competitors have snatched market share. Therefore, designers are under pressure to design chips as quickly as possible. Design time is especially tight in application-specific IC design, where only a few weeks may be available to turn a concept into a working ASIC.
Designers have developed two techniques to eliminate unnecessary detail: hierarchical design and design abstraction. Designers also make liberal use of computer-aided design tools to analyze and synthesize the design.
1.4.1 Hierarchical Design
Divide-and-conquer
Hierarchical design is commonly used in programming: a procedure is written not as a huge list of primitive statements but as calls to simpler procedures. Each procedure breaks down the task into smaller operations until each step is refined into a procedure simple enough to be written directly. This technique is commonly known as divide-and-conquer—the procedure’s complexity is conquered by recursively breaking it down into manageable pieces.
Components
Chip designers divide and conquer by breaking the chip into a hierarchy of components. As shown in Figure 1-4, a component consists of a body and a number of pins—this full adder has pins a, b, cin, cout, and sum. If we consider this full adder the definition of a type, we can make many instances of this type. Repeating commonly used components is very useful, for example, in building an n-bit adder from n full adders. We typically give each component instance a name. Since all components of the same type have the same pins, we refer to the pins on a particular component by giving the component instance name and pin name together; separating the instance and pin names by a dot is common practice. If we have two full adders, add1 and add2, we can refer to add1.sum and add2.sum as distinct terminals (where a terminal is a component-pin pair).
)
Figure 1-4 Pins on a component.
Net lists
We can list the electrical connections which make up a circuit in either of two equivalent ways: a net list or a component list. A net list gives, for each net, the terminals connected to that net. Here is a net list for the top component of Figure 1-5:
net1: top.in1 i1.in net2: i1.out xxx.B topin1: top.n1 xxx.xin1 topin2: top.n2 xxx.xin2 botin1: top.n3 xxx.xin3 net3: xxx.out i2.in outnet: i2.out top.out
)
Figure 1-5 A hierarchical logic design.
A component list gives, for each component, the net attached to each pin. Here is a component list version of the same circuit:
top: in1=net1 n1=topin1 n2=topin2 n3=topin3 out=outnet i1: in=net1 out=net2 xxx: xin1=topin1 xin2=topin2 xin3=botin1 B=net2 out=net3 i2: in=net3 out=outnet
Given one form of connectivity description, we can always transform it into the other form. Which format is used depends on the application—some searches are best performed net-by-net and others component-by-component. As an abuse of terminology, any file which describes electrical connectivity is usually called a netlist file, even if it is in component list format.
As shown in Figure 1-5, a logic design can be recursively broken into components, each of which is composed of smaller components until the design is described in terms of logic gates and transistors. In this figure, we have shown the type and instance as instance(type); there are two components of type A. Component ownership forms a hierarchy. The component hierarchy of Figure 1-5 is shown in Figure 1-6. Each rounded box represents a component; an arrow from one box to another shows that the component pointed to is an element in the component which points to it. We may need to refer to several instance names to differentiate components. In this case, we may refer to either top/i1 or top/i2, where we trace the component ownership from the most highest-level component and separate component names by slashes (/). (The resemblance of this naming scheme to UNIX file names is intentional—many design tools use files and directories to model component hierarchies.)
)
Figure 1-6 A component hierarchy.
Components as black boxes
Each component is used as a black box—to understand how the system works, we only have to know each component’s input-output behavior, not how that behavior is implemented inside the box. To design each black box, we build it out of smaller, simpler black boxes. The internals of each type define its behavior in terms of the components used to build it. If we know the behavior of our primitive components, such as transistors, we can infer the behavior of any hierarchically-described component.
People can much more easily understand a 100,000,000-transistor hierarchical design than the same design expressed directly as ten million transistors wired together. The hierarchical design helps you organize your thinking—the hierarchy organizes the function of a large number of transistors into a particular, easy-to-summarize function. Hierarchical design also makes it easier to reuse pieces of chips, either by modifying an old design to perform added functions or by using one component for a new purpose.
1.4.2 Design Abstraction
Levels of modeling
Design abstraction is critical to hardware system design. Hardware designers use multiple levels of design abstraction to manage the design process and ensure that they meet major design goals, such as speed and power consumption. The simplest example of a design abstraction is the logic gate. A logic gate is a simplification of the nonlinear circuit used to build the gate: the logic gate accepts binary Boolean values. Some design tasks, such as accurate delay calculation, are hard or impossible when cast in terms of logic gates. However, other design tasks, such as logic optimization, are too cumbersome to be done on the circuit. We choose the design abstraction that is best suited to the design task.
We may also use higher abstractions to make first-cut decisions that are later refined using more detailed models: we often, for example, optimize logic using simple delay calculations, then refine the logic design using detailed circuit information. Design abstraction and hierarchical design aren’t the same thing. A design hierarchy uses components at the same level of abstraction—an architecture built from Boolean logic functions, for example—and each level of the hierarchy adds complexity by adding components. The number of components may not change as it is recast to a lower level of abstraction—the added complexity comes from the more sophisticated behavior of those components.
The next example illustrates the large number of abstractions we can create for a very simple circuit.
Example 1-2. Layout and its abstractions
Layout is the lowest level of design abstraction for VLSI. The layout is sent directly to manufacturing to guide the patterning of the circuits. The configuration of rectangles in the layout determines the circuit topology and the characteristics of the components. However, the layout of even a simple circuit is sufficiently complex that we want to introduce more abstract representations that help us concentrate on certain key details.
Here is a layout for a simple circuit known as a dynamic latch:
)
This layout contains rectangles that define the transistors, wires, and vias which connect the wires. The rectangles are drawn on several different layers corresponding to distinct layers of material or process steps in the integrated circuit.
Here is an abstraction for that layout: a stick diagram, which is a sketch of a layout:
)
This stick diagram has the same basic structure as the layout, but the rectangles in the layout are abstracted here as lines. Different line styles represent different layers of material: metal, diffusion, etc. Transistors are formed at the intersection a line representing polysilicon with either a n-type or p-type diffusion line. The heavy dots represent vias, which connect material on two different layers. This abstraction conveys some physical information but not as much as the layout—the stick diagram reflects the relative positions of components, but not their absolute positions or their sizes.
Going one more step up the abstraction hierarchy, we can draw a transistor-level schematic:
)
This formulation is not intended to describe the physical layout of the circuit at all—though the placement of transistors may resemble the organization of the transistors in the layout, that is a matter of convenience. The intent of the schematic is to describe the major electrical components and their interconnections.
We can go one step higher in the abstraction hierarchy to draw a mixed schematic:
)
This is called mixed because it is built from components at different levels of abstraction: not only transistors, but also an inverter, which is in turn built from transistors. The added abstraction of the inverter helps to clarify the organization of the circuit.
The next example shows how a slightly more complex hardware design is built up from circuit to complex logic.
Example 1-3. Digital logic abstractions
A transistor circuit for an inverter is relatively small. We can determine its behavior over time, representing input and output values as continuous voltages to accurately determine its delay:
)
We can use transistors to build more complex functions like the full adder. At this point, we often simplify the circuit behavior to 0 and 1 values which may be delayed in continuous time:
)
As circuits get bigger, it becomes harder to figure out their continuous time behavior. However, by making reasonable assumptions, we can determine approximate delays through circuits like adders. Since we are interested in the delay through adders, the ability to make simplifying assumptions and calculate reasonable delay estimates is very important.
When designing large register-transfer systems, such as data paths, we may abstract one more level to generic adders:
)
){kind=link}
At this point, since we don’t know how the adders are built, we don’t have any delay information. These components are pure combinational elements—they produce an output value given an input value. The adder abstraction helps us concentrate on the proper function before we worry about the details of performance.
Design abstractions
Figure 1-7 shows a typical design abstraction ladder for digital systems:
- Specification. The customer specifies what the chip should do, how fast it should run, etc. A specification is almost always incomplete—it is a set of requirements, not a design.
- Behavior. The behavioral description is much more precise than the specification. Specifications are usually written in English, while behavior is generally modeled as some sort of executable program.
- Register-transfer. The system’s time behavior is fully-specified—we know the allowed input and output values on every clock cycle—but the logic isn’t specified as gates. The system is specified as Boolean functions stored in abstract memory elements. Only the vaguest delay and area estimates can be made from the Boolean logic functions.
- Logic. The system is designed in terms of Boolean logic gates, latches, and flip-flops. We know a lot about the structure of the system but still cannot make extremely accurate delay calculations.
- Circuit. The system is implemented as transistors.
- Layout. The final design for fabrication. Parasitic resistance and capacitance can be extracted from the layout to add to the circuit description for more accurate simulation.
)
Figure 1-7 A hierarchy of design abstractions for integrated circuits.
Top-down and bottom-up design
Design always requires working down from the top of the abstraction hierarchy and up from the least abstract description. Obviously, work must begin by adding detail to the abstraction—top-down design adds functional detail. But top-down design decisions are made with limited information: there may be several alternative designs at each level of abstraction; we want to choose the candidate which best fits our speed, area, and power requirements. We often cannot accurately judge those costs until we have an initial design. Bottom-up analysis and design percolates cost information back to higher-levels of abstraction; for instance, we may use more accurate delay information from the circuit design to redesign the logic. Experience will help you judge costs before you complete the implementation, but most designs require cycles of top-down design followed by bottom-up redesign.
1.4.3 Computer-Aided Design
CAD tools
The only realistic way to design chips given performance and design time constraints is to automate the design process, using computer-aided design (CAD) tools which automate parts of the design process. Using computers to automate design, when done correctly, actually helps us solve all three problems: dealing with multiple levels of abstraction is easier when you are not absorbed in the details of a particular design step; computer programs, because they are more methodical, can do a better job of analyzing cost trade-offs; and, when given a well-defined task, computers can work much more quickly than humans.
Design entry
Computer-aided design tools can be categorized by the design task they handle. The simplest of CAD tool handles design entry—for example, an interactive schematic drawing package. Design entry tools capture a design in machine-readable form for use by other programs, and they often allow easier modification of a design, but they don’t do any real design work.
Analysis and verification
Analysis and verification tools are more powerful. The Spice circuit simulator, for example, solves the differential equations which govern how the circuit responds to an input waveform over time. Such a program doesn’t tell us how to change the circuit to make it do what we want, but many analysis tasks are too difficult to perform manually.
Synthesis
Synthesis tools actually create a design at a lower level of abstraction from a higher level description. Some layout synthesis programs can synthesize a layout from a circuit description like that in Figure 1-2. Using computers for design is not a panacea. Computer programs cannot now, nor are they ever likely to be able to transform marketing brochures directly into finished IC designs. Designers will always be necessary to find creative designs and to perform design tasks which are too subtle to be left to algorithms.
Both hierarchical design and design abstraction are as important to CAD tools as they are to humans—the most powerful synthesis and analysis tools operate on a very restricted design model. CAD tools can help us immensely with pieces of the design task, but algorithms that have the detailed knowledge required to solve one design problem usually do not have the broad range of data required to balance broad requirements.
Tools as aids
CAD tools must be used judiciously by a human designer to be most effective. Nonetheless, CAD tools are an essential part of the future of IC design because they are the only way to manage the complexity of designing large integrated circuits. Manual design of a hundred-million transistor chip, or even a 100,000 transistor chip, quickly overwhelms the designer with decisions. Not all decisions are equally important—some may have only a minor effect on chip size and speed while others may profoundly change the chip’s costs. By concentrating on the wrong decisions, a designer may cause problems that are not easily correctable later. CAD tools, by automating parts of the design process, help the designer eliminate mundane decisions quickly and concentrate on the make-or-break problems posed by the chip.
For example, long wires can introduce excessive delay, increase power consumption, and create opportunities for crosstalk. Such problems can be found by a program that analyzes delays through the chip, but when designing a chip by hand, it may be easy to miss this single connection, and the error will not be found until the chip comes back from fabrication. CAD tools are particularly important for evaluating complex situations in which solving one problem creates other problems—for example, making one wire shorter makes other wires longer. When two constraints compete, solutions to problems may not be so easy. Making one part of the design faster may, for example, make another part of the design unacceptably large and slow. CAD tools help us solve these problems with analytical methods to evaluate the cost of decisions and synthesis methods that let us quickly construct a candidate solution to a problem. Evaluation of candidate designs is critical to designing systems to satisfy multiple costs because optimizing a complete system cannot be done simply by optimizing all the parts individually—making each part in a chip run as fast as possible in isolation by no means ensures that the entire chip will run as fast as possible. Using CAD tools to propose and analyze solutions to problems lets us examine much larger problems than is possible by hand.