What is a chiplet, and why split one big chip into several small ones?

A chiplet is what you get when a chip that would have been made as one big chip is split into several smaller chips. Each one is made and tested on its own; then they are placed side by side or stacked, put into one package (the casing around a chip that protects it and connects it to the outside), and work like a single chip. The "-let" in chiplet means small. A chip made all in one piece is called monolithic. Much of what the news calls "advanced packaging" is about putting chiplets back together.

Why split something apart, only to put it back together? Let's start from the beginning.

Why split one big chip into several?

There are three main reasons.

One defect ruins only one small piece

A wafer always picks up a few defects too small to see, and whichever chip one lands on is ruined. The bigger the chip, the more likely it is to be hit, and the lower the yield (What does "yield" mean in chip news? explains this with cookies).

If a single defect lands on a big chip, the whole chip often has to be thrown away. Split it into four small chips, and the same defect hits only one of them: you throw away that one small piece, and the other three are still good. They can be matched up with good chiplets from elsewhere to make a full set.

Big chips also have a size limit: a scanner prints one field at a time, so a chip can hardly be made bigger than one field, and fitting in more usually means splitting it up and putting it back together (The news says "advanced packaging" covers this).

Each piece is made on the generation that suits it

One chip can only be made with one generation of manufacturing. Once it is split into chiplets, the computing one can use the newest generation, while the ones that talk to the outside world or handle analog signals, which hardly get any smaller on a newer generation, use older generations that are cheaper and well proven, so the money goes where it matters most (The news says "advanced packaging" covers this; for older generations, see The news says "mature nodes").

The same chiplet can build different products

Chiplets work like building blocks. The same computing chiplet, used a few times or many times, makes smaller or bigger products; the chiplet that talks to the outside world can be shared by several products.

Designing a chip and making a set of photomasks both cost a lot (Why are chips so expensive?). The more products one chiplet goes into, the less each product has to carry; and a chiplet that has already been made and proven can go straight into the next product.

Step by step: how is a chip split, and put back together?

Step 1: Cut the design into pieces

• What: While the chip is being designed, decide which functions go together on one chiplet, for example one for computing and one for talking to the outside world.
• Why: Every cut turns a stretch of wiring inside the chip into a road between two chiplets. The best place to cut is where little data crosses, or where it isn't in a hurry.
• Otherwise: Cut where data goes back and forth the most, and the road jams: the whole chip slows down and uses more power.

Step 2: Make each piece, and test each one first

• What: Each piece is made separately in a wafer fab. While it's still on the wafer, every chip is tested by wafer sort with a probe card, and only the good ones move on (How are chips tested?).
• Why: If even one chip in a package is bad, the whole set often can't be used, and the good ones go to waste with it.
• Otherwise: A bad one slips in and is found only after assembly, and the good chiplets and the packaging cost are lost along with it.

The contacts between chiplets are tiny and packed close together, so the needles can't always reach every one. That's why designers often add extra contacts just for the needles, or build circuits into the chip that help it check itself.

Step 3: Place the good ones side by side or stack them, with dense connections

• What: The good chiplets are placed side by side on a thin slice with very fine wiring, or stacked on top of each other, and connected by many very fine, very short wires.
• Why: Chiplets pass data to each other all the time. The shorter the wires and the more of them, the faster and more efficiently the data moves, so the set works like one chip.
• Otherwise: Package each one separately and connect them through a circuit board, the traditional way, and the road is long with few lanes. The money saved by splitting is eaten up by the lost speed.

The densest way needs no little solder bumps at all: the copper contacts on the surfaces of two chips, together with the insulating layer around them, are lined up and bonded directly, copper to copper. This is called hybrid bonding. For how stacked chips connect from top to bottom, see "What is a TSV (through-silicon via)?".

Step 4: Test the whole set again

• What: Once packaged, the whole set is tested again as one chip; this is the final test.
• Why: While stacking, soldering and bonding, a contact may not connect, or one chiplet may get damaged. And whether the road between chiplets works can only be tested once they are put together.
• Otherwise: A faulty set goes into a product and fails in the user's hands. Once packaged, taking the bad chiplet out and replacing it is hard, which is why step 2 has to catch it first.

What does splitting cost?

• A slightly longer road: inside one chip, a signal just travels along the chip's own wiring. After the split, it has to leave one chiplet, travel through the wiring in the package, and enter the other. However short and dense that road is, it is a little slower than staying inside one chip, and each 0 or 1 sent takes a little more power.
• Extra interface circuits: every chiplet needs an extra block whose job is to send and receive data, one on each side of every link. Added up, the silicon and the power used are more than a single chip would need.
• Harder, more expensive packaging: the wires have to be very dense and lined up very precisely; with several chips packed close together, heat is harder to get out; and the contacts are tiny and packed close, so testing each chiplet first takes more work.

So not every chip is worth splitting. If a chip isn't big and doesn't need to mix generations, making it as one chip is simpler and cheaper. For how chiplets differ from a system on a chip (SoC) and from SiP, see the table in "What is an SoC?".

How can chiplets made by different people connect?

Before two chiplets can pass data to each other, both sides have to agree on how to connect: how the contacts are laid out, how many wires there are, how signals are sent, and what format the data is in. It's like a plug and a socket: they have to be the same shape before an appliance can plug in.

Chiplets that one company makes for itself only need that company's own agreement. But for chiplets designed by different people, or even made in different fabs, to work together, everyone needs to agree on one common interface. The industry has been pushing for common interfaces like this for several years. Once they are settled, building a big chip may become a bit like picking parts: choose a few ready-made chiplets, and design only the piece you do best.

In fab24hr

There are no chiplets in the game. In the main story, The life of a chip, the hearing-aid chip Glowmere is a single chip. In act three, at the test-and-assembly house, it goes into an SiP module: a small board with this one chip and a few resistors and capacitors, connected by wire bonding.

Still, "test first, package only the good ones" is something you can see in the game. In chapter 10, "Standby Current", you're the test engineer: every chip on Glowmere's first production wafers has to be tested, and you look at the bins and the wafer map before deciding where to set the standby current limit. In chapter 11, "The Ball Lifted", the wafers have been cut, and only the good chips go into modules; tap the die bonder on the assembly line and its card says that chips that failed wafer test are not picked. In "A day in the fab", as the assembly process engineer at the test-and-assembly house, the morning shift, "Chipping", also asks you to decide whether chips with a small nick at the edge are kept.

The companies, chips and people in the game are fictional; what chiplets are and why chips are split follow public, general sources. For other jargon, see Chip fab jargon in plain English.