what is a chiplet
A chiplet is a small, specialized piece of silicon designed to be combined with other chiplets inside a single chip package, rather than building one enormous monolithic die that does everything — an approach that lets chipmakers mix and match specialized pieces, improve manufacturing yield, and update parts of a design independently.
```flowchart
{
"rows": [
{ "type": "nodes", "items": [
{ "title": "Traditional approach: one large monolithic die", "sub": "everything built on a single piece of silicon", "tone": "blue" }
]},
{ "type": "arrow" },
{ "type": "group", "title": "Chiplet approach instead", "items": [
{ "title": "Several smaller specialized dies", "sub": "e.g. compute, memory, I/O as separate chiplets", "tone": "green" },
{ "title": "Connected via advanced packaging", "sub": "short, high-bandwidth links inside one package", "tone": "green" }
]},
{ "type": "arrow" },
{ "type": "nodes", "items": [
{ "title": "Behaves as one chip to the outside world", "sub": "better yield and flexibility than a single giant die", "tone": "orange" }
]}
]
}
```
**Chiplets exist largely to solve a manufacturing yield problem that gets worse as a single die gets larger.** A larger monolithic die has a higher chance of containing at least one manufacturing defect somewhere on it, which can ruin the entire die; splitting the same overall design into several smaller chiplets means a single defect only ruins one small chiplet rather than an entire large die, meaningfully improving how many usable pieces come out of a given amount of wafer production.
```svg
```
| Approach | Yield impact | Design flexibility |
|---|---|---|
| Monolithic die | Larger die = more exposure per defect | Everything must be redesigned together |
| Chiplet-based design | Smaller pieces = defects affect less | Individual chiplets can be updated or swapped independently |
**Chiplets also let a chip design mix and match manufacturing processes optimized for each specific function.** A compute chiplet might benefit from the newest, most expensive process node, while a memory or input/output chiplet might work perfectly well, and far more cheaply, on an older, more mature process — building both as separate chiplets on their own most appropriate process lets a chipmaker avoid the cost of manufacturing an entire large die on the single most expensive process available.
**Making chiplets work well together depends heavily on the connections between them, which is why advanced packaging and chiplets have developed together.** The short, high-bandwidth links needed to make several separate chiplets function together as if they were one chip require the same kind of advanced packaging techniques used for stacking dies — without those tightly engineered interconnects, splitting a design into chiplets would introduce too much delay and power loss between the pieces to be worthwhile.
**Chiplets also give chipmakers a path to reuse proven designs across multiple products, rather than redesigning everything from scratch each time.** A chiplet built for one product's compute needs can potentially be reused, paired with different companion chiplets, across a whole family of products, letting a design team amortize the cost of designing and validating that one chiplet over many different final chip configurations.
Read the chiplet through a modular-building-blocks lens: instead of designing and manufacturing one enormous piece of silicon that has to work perfectly all at once, chiplets let chipmakers build smaller specialized pieces that fail less often individually and can be mixed, matched, and reused far more flexibly than a single monolithic die ever could.