How does a chip design company work with a foundry, from order to delivery?

When the news says a chip "has started production at a foundry", there's a long process behind it between a chip design company and a wafer fab: choose a process and get the design rules, book capacity and dates, hand over the design to be made into photomasks, start the wafers and build them in the fab for months, check them, then send them to a test-and-assembly house to be tested and packaged. From tape-out to chips in hand often takes several months.

What a foundry is was covered in The news says "foundry". What is it?; this article is about how the two companies work together, step by step.

What are the steps from order to delivery?

Think of publishing a book. The publisher hands the manuscript to a printer, they agree on the paper and the print date, then the printer makes the printing plates, prints and binds. A chip works the same way, except every step takes longer and costs more. There are roughly eight steps.

Step 1: Choose a process, get the design rules and device models

• What: Pick a foundry, a process generation, and one version within that generation (a generation often comes in a faster version, a lower-power version and one with especially low leakage on standby). The foundry hands over a set of design files, often called a process design kit (PDK): the design rules (how thin a line can be, how close two lines can get), the transistor models (which a computer uses to work out how fast a circuit is and how much power it uses) and the rule files for checking.
• Why: Transistors come out differently at every foundry and in every generation, so the design has to be drawn for what this fab can actually make (see How are chips designed?).
• Otherwise: With the wrong rules, the design fails its checks; with the wrong models, the finished chip's speed and power use come out far from what the computer predicted.

Step 2: Agree on capacity and dates

• What: The two sides agree roughly when the design will tape out, how many wafers go in first and how many a month after that; the foundry checks whether its tools can fit it in and holds the slots. Some contracts also have the customer pay up front to reserve capacity.
• Why: A wafer takes months to make, and the tools are often fully booked (see What does production control do in a chip fab?).
• Otherwise: The design is finished but can't get in, so it waits; and if the product sells well, there's no way to make more.

Step 3: Tape out, hand over the design files

• What: Once every check on the design has passed, the final layout files (the shapes on every layer) go to the foundry. This is called tape-out (see What is tape-out?). The foundry checks them again itself when they arrive.
• Why: The next step, making masks, is expensive and slow, so mistakes have to be caught here.
• Otherwise: A mistake found only after the wafers are made costs money and several months.

Step 4: Make the photomasks

• What: The patterns for the finest layers are first corrected by computer; then each layer is made into a photomask and inspected. A photomask is the "negative" used to print the circuit, and a full set often has dozens (see What is a photomask?).
• Why: The pattern is printed onto the wafer with light shining through the mask, just as printing needs a plate first.
• Otherwise: One defect missed on a mask gets printed onto every wafer.

Step 5: Wafer start

• What: Once the masks have arrived and the route is set, the first lot of wafers goes onto the line and starts its first step. This is the wafer start (see What is a wafer start?).
• Why: How much can be delivered months later, and roughly when, is mostly settled on this day.
• Otherwise: Start a week late and you ship a week late; adding more partway through usually means starting again from the first step.

Step 6: The wafers go step by step, for months

• What: The wafers go through hundreds of steps, building up layer by layer; people in the industry often say about three months on average, longer for more advanced processes (see How long does a wafer take to make?). When a tool stops along the way or a lot is put on hold, the foundry has to tell the customer whether the lot will be late.
• Why: A chip has dozens of layers, and each one has to be printed, etched, given a new film and measured; most of the time, the wafers are waiting in line for tools.
• Otherwise: Skip a step and the chip is missing a layer, so the only way to go faster is to wait less (for example, as a hot lot), not to do less.

Step 7: Measure WAT before the wafers leave, hand over the data

• What: When the wafers are finished, WAT comes first: probes measure small structures left between the chips just for measuring, to check basic numbers such as transistors and resistance (see What is WAT?). The numbers often go to the customer along with the wafers.
• Why: They show that this lot was made with the agreed process and matches the models the customer designed with.
• Otherwise: If the lot drifted and nobody noticed, it gets tested, cut and packaged anyway, and the money and time are wasted.

Step 8: Send the wafers to the test-and-assembly house

• What: The wafers go into sealed shipping boxes and travel, as the customer has arranged, to the assembly and test house. There, wafer sort tests every chip and draws a wafer map; then the chips are cut apart, packaged and tested again (see How are chips tested?).
• Why: The chips are still joined together on the wafer and have no package, so they can't be soldered onto a circuit board yet.
• Otherwise: Bad chips that weren't caught get packaged along with the good ones, which costs more and may even reach the customer.

Why do first tries often "share a ride"?

Making a whole mask set just for a few sample chips is very expensive. A common way around it is a multi-project wafer, MPW for short, also called a shuttle: designs from several companies are squeezed onto the same mask set and made on one batch of wafers, and the cost is shared. It's like taking a shuttle bus: you don't hire the whole bus, but it leaves on a timetable.

• The upside: For a small share of the cost, you get real chips made in a fab. Universities use it too, so students can have chips they designed made for real.
• The cost: The departure dates are fixed, and if you miss one you wait for the next. The number of chips you get back is limited, mostly for testing; a product made in large numbers still needs its own full mask set.

What's the difference between an engineering lot and a production lot?

Even with its own full mask set, a new product's first few lots are usually engineering lots: made to try things out and find problems, not to sell. Some wafers are often made to run deliberately fast or slow, to see if the design holds up. That's a split lot, and the ways the wafers are skewed are called corners.

A production lot is made to go to the customer and be sold. It follows a locked standard process, the POR, and every lot has to come out the same. Before moving to production there are usually a few gates: yield is steady enough, the reliability tests have passed, and the customer has signed off. Starting before the tests have finished is called a risk start.

What if it comes out wrong, or something has to change?

The first lot is often not right first time. The design company puts the chips on a board and tests them, and the foundry looks at the WAT and the wafer maps to find which step drifted; working from the same data, the two find the causes, and yield climbs lot by lot (see Why do bigger chips have lower yield?).

• The design is wrong: change the design and tape out again, remake the masks that changed, and start the wafers from the beginning. A common way to save money is to scatter some spare logic gates through the chip ahead of time; a fix then only touches the wiring above the transistors, and only the masks for the layers that changed are remade.
• The foundry wants to change its process: the customer qualified the chip as it was originally made, so a change that affects the customer has to be formally announced to them first (see What is change control in a chip fab?).

Next time you see it in the news

• "The new chip has taped out": the design has been handed over and the masks are about to be made. Nobody has chips yet.
• "Production has started at a foundry": the wafers are being made. That takes months, and testing and packaging still come after.
• "Samples are shipping": some chips are being sent to the companies that will buy them, to try out and check; it isn't volume shipping yet.
• "Orders are booked into next year": most of the capacity until next year already has customers' orders in it, and new orders have to wait for a slot to open.

In fab24hr

The main story, The life of a chip, follows a made-up hearing-aid chip, "Glowmere". In the prologue, "Tape-out", its design 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, deciding how to split the first 12 wafers, and the customer's product engineer, Leo, video-calls to look at the split plan. In chapter 4, "Two Weeks Early", the reliability test still has three weeks to run, and you advise the customer whether to wait for it or go ahead with a risk start.

Tap "Clock in", choose "A day in the fab", then on the "Your role today" card tap "Production fab" and swipe to the process integration engineer (marked Beta). Late at night, the electrical results for a new product's first lot come in, and it's due to ship to the test-and-assembly house at eight in the morning; in the evening, the customer, Leo, is on the line waiting for your explanation. Tap "Test-and-assembly house" and swipe to the customer liaison: the evening shift is chapter 12 of the main story, "The Night Before Shipping".

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