How are chips designed, and what does IC design involve?

In one sentence: a chip is designed on a computer and checked many times before a fab ever makes it. IC design (IC stands for integrated circuit, the formal name for a chip) is that first part: deciding what the chip should do, writing the circuit, making sure it works, laying it out as patterns layer by layer, and handing it over to be made into photomasks.

How design jobs differ from fab jobs is covered in "VLSI design vs working in a fab: what's the difference?". This article walks through the design itself, step by step.

What are the steps in designing a chip?

It's like building a house: decide what you need, draw the plans, check them, and start building only once they're approved. Designing a digital chip goes roughly like this:

1. Spec: what the chip does, how fast, how little power.
2. Write the behavior: describe what the circuit does in a hardware description language.
3. Verify: make sure it does what the spec says.
4. Synthesize: turn it into logic gates.
5. Place and route: decide where each gate goes and how the wires connect.
6. Check: can it be made, and is it fast enough?
7. Tape out: hand it to the fab to make the masks.

In practice, teams go back and redo steps many times.

What goes into the spec?

The spec is the chip's list of requirements: what it does, how fast it runs, the most power it may use, and how big it is. Size affects price: the bigger the chip, the fewer fit on one wafer, so each one costs more.

These pull against each other. Going faster often takes more circuitry and more power; saving power may mean running a bit slower. For something that has to last a long time on one small battery, power comes first.

How do you "write" a circuit?

A phone chip holds more than ten billion transistors (What is a transistor?), far too many to draw one by one. So digital circuits are usually written as text, in a hardware description language (HDL).

It looks a bit like programming. The difference: a program runs one line at a time, in order; an HDL describes a circuit, and many parts of it work at the same time.

Most digital circuits move to a steady beat called the clock, like a metronome. A common way to write them spells out, for every beat, where data comes from, what is calculated and where it is stored. This style is often called RTL.

Why does verification take so much effort?

Verification means making sure the circuit you wrote does what the spec says. The general view is that it often takes a large share of design work, because:

• You can't try everything: the inputs, plus everything the circuit is holding in memory, add up to far too many cases to test one by one.
• Bugs hide in rare cases: everything works until two things happen at exactly the same moment.
• Missing one is expensive: find it after tape-out, and you have to change the design and make new masks.

Three common approaches:

• Simulation: give the circuit all kinds of inputs on a computer and check the outputs, like setting it lots of exam questions.
• Formal verification: instead of setting questions, prove with math that something is always true, for example "these two signals are never switched on together".
• Try it out: load the design into a chip whose circuit can be rewritten (an FPGA) and actually run it.

How does the written behavior become a circuit?

This step is synthesis: design software reads the behavior and translates it into a circuit of connected logic gates. A logic gate is a tiny circuit of a few transistors, such as "output 1 only if both inputs are 1" (How is memory different from a processor?).

It builds from a cell library: small blocks designed and checked in advance, including all kinds of logic gates and tiny circuits that remember a 0 or a 1. Every block is the same height but a different width, so they line up in rows, like building blocks. The same gate has a different size and speed in each generation of process, so a library is made for one fab's particular process.

Where does everything go, and how is it wired?

Synthesis produces a list of what connects to what, with no positions yet. Place and route turns it into real patterns:

1. Floorplan: where the big blocks go and where power comes in.
2. Placement: put the gates in, keeping ones that connect a lot close together.
3. Clock tree: get the beat to every part of the chip at nearly the same time.
4. Routing: several layers of wiring are stacked like a multi-level highway interchange, and the wires run through them.

The result is one pattern per layer, called the layout. Later, each layer gets at least one mask.

What gets checked before it's handed over?

Everything has to pass before it can go out. This stage is often called sign-off:

• Can it be made? The fab has a set of design rules, such as how thin a line can be and how close two lines can get. Design software checks the layout against every rule: a design rule check (DRC).
• Is the drawing the same circuit as the code? Work the circuit back out of the layout and compare it with the original: LVS (layout versus schematic).
• Is it fast enough? On every beat, each signal has to reach its next stop before the next beat; the slowest path sets how fast the chip can run. Transistors from the fab come out a little fast or a little slow, and these ways of being off are called corners: the chip has to keep up at the slow ones, and it can't go wrong at the fast ones either.
• Can it be tested later? Test circuitry is added in advance, so a tester can look inside every chip once it's made (How are chips tested?).

What is tape-out?

Once everything passes, the final design files go to the fab. This is tape-out. The usual explanation is that the final files were once handed over on paper tape and later magnetic tape, and the name stuck.

Changing the design after tape-out means making new masks: expensive, and months of extra waiting. And tape-out isn't the end. Once the chips are made, they still need testing to see if they really do what the spec says; if not, the design is changed and taped out again. For how tape-out differs from a wafer start, see "What is a wafer start?".

Are analog circuits designed the same way?

Not quite. Analog circuits handle signals that change smoothly, such as sound, voltage and radio (Analog vs digital chips: what's the difference?). Their design is usually more hands-on:

• Engineers often decide how big each transistor is and how it connects themselves, then tune it again and again in circuit simulation.
• The layout is often drawn by hand too. Analog circuits are sensitive to placement: two transistors that must match are laid out symmetrically and close together, and a noisy circuit next door can interfere.

Tools that automate this are improving, but the work still leans heavily on experience.

How do designers and fabs work together?

A design has to fit what the fab can make, so the two sides connect from the start:

• Before design starts: the fab provides a set of design files, often called a process design kit (PDK): the design rules, transistor models (usually the typical case and several corners) and rule files for the checks. Sometimes it includes a cell library; sometimes someone else builds one for that process.
• After tape-out: the masks are made, and the first wafers go in (the wafer start). A new chip's first lots are a pilot run: designers look at the test results, and the fab looks for which step caused a problem.

Design and manufacturing are often done by different companies (The news says "fabless", "foundry" and "IDM". What's the difference?); a company that designs and makes its own chips follows the same kind of rules.

In fab24hr

There's no chip-designer role in the game; the main story, The life of a chip, starts after tape-out. The first time you play it, you see the prologue, "Tape-out": "Glowmere", the first chip a small hearing-aid company has designed itself, goes to the fab and becomes a set of photomasks.

Chapter 1, "How to Split 12", is the day before wafer start. You're the R&D integration engineer on the R&D line: the same design never comes out exactly the same, so a few of the first 12 wafers are made fast or slow on purpose to see if the design holds up (What does an R&D integration engineer do with a new chip's first lot?). In chapter 3, "The shape of a failure", the next mask set is finalized next week, and you decide whether to leave room inside the chip for your own test keys, though the area belongs to the customer.

The companies, chip and people in the game are fictional; the design flow comes from public, general sources, and every company does it differently. For other jargon, see Chip fab jargon in plain English.