UCIe · Module 1

The Chiplet Revolution

Why splitting a system across dies was not the real shift — and how designing dies as reusable, composable building blocks changed what a product is, where process choice is made, and why die-to-die interfaces had to become published contracts.

Multi-Die Systems showed how one logical system is partitioned across several dies and what every cut costs. It would be easy to stop there and call that the chiplet revolution. It is not. A company can split its own product across its own dies, connect them with an interface of its own invention, and ship something genuinely multi-die that is no more modular than the monolithic design it replaced — each die still exists for exactly one product, and nothing about it is reusable. This chapter is about the step that actually mattered: when dies stopped being pieces of one product and started being building blocks.

1. Decomposition Was Not the Revolution

Hold the two ideas apart, because they get collapsed constantly.

Decomposition answers how do we build a system too large, too expensive, or too coupled to sit on one die? You cut it up. Chapter 1.6 covered that completely, and it is a real engineering achievement.

Composition answers a different question: how do we build many products out of a small number of known pieces? That requires the pieces to be worth keeping — designed, verified, and manufactured once, then used repeatedly.

The revolution is the shift from product-specific dies to reusable, composable dies.

A useful test: if a die exists only inside one product and would be redesigned for the next one, decomposition happened. If the same manufactured die appears in several products and the products differ by which pieces they combine, composition happened. Both are multi-die. Only the second changes what a product fundamentally is.

2. Four Stages

The industry did not jump from one to the other. It moved through recognisable stages, and knowing which one a given design sits in is more useful than the label on the box.

Figure 1 — the progression from one inseparable die to composable building blocks. The critical change is the bottom row: what unit of work gets reused. Decomposition (Stage 2) still rebuilds a die per product; only when the die itself becomes the reusable unit (Stage 3) does product differentiation become recombination rather than redesign. Stage 4 is a direction the industry is moving in, not a description of how most silicon ships today.

Stage 1 — Monolithic. Everything on one die. Reuse exists, but only as RTL: the design is carried forward and the silicon is rebuilt every time.

Stage 2 — Proprietary multi-die. One organisation splits one product across several dies, joined by an internal link it designed. This solves size, yield, and closure pressure — but the dies were drawn for this product, and the next product gets new ones. The reuse unit has not moved.

Stage 3 — Reusable chiplets. Dies are designed with a second question in mind: what else could use this? The interface is stabilised so the die can meet different partners, and one manufactured design serves several products. This is where the reuse unit finally moves from the design to the silicon.

Stage 4 — Ecosystem composition. The stated ambition: dies from different companies, designed independently, combined into one package because they share an agreed interface. Treat this as a direction the industry is working toward rather than a description of how most silicon ships today — the technical groundwork is being laid, and broad plug-and-play sourcing of third-party dies is not the normal case.

3. What Turns a Die Into a Building Block

A full definition belongs to the next chapter. What matters here is the intent that separates Stage 3 from Stage 2, because it is a design-time decision rather than a property discovered afterwards.

A die behaves like a building block when it is meant to be a functional subsystem rather than an arbitrary slice, when it is independently manufactured and tested as its own object, when it connects through a defined die-to-die interface rather than an implicit private one, and — the decisive clause — when it is built with reuse or composition in mind beyond the first product.

That last point is what makes this an architectural choice rather than a manufacturing accident. Two dies can be physically identical in every respect except that one was drawn to fit exactly one neighbour and the other was drawn to meet several. Only the second is a building block, and the difference lives in the interface and in who was consulted when it was specified.

4. Reuse Changes What a Product Is

Chapter 1.5 established that a monolithic die makes the die the unit of reuse, so RTL reuse still means rebuilding silicon. Stage 3 is where that finally breaks.

If a compute die can appear in a small, a medium, and a large product, one physical design is amortised across all three. If an I/O die can serve several different compute configurations, the same holds from the other direction. The consequence is a change in how products are differentiated:

Product differentiation moves from redesigning a die to recombining known dies.

That is a genuinely different development model — but it is not free, and this is where enthusiasm usually overshoots. Every combination still needs integration work, a package designed for that arrangement, validation of that specific combination, and ongoing compatibility management as the pieces evolve independently. What reuse removes is redesigning and re-closing the die. What it does not remove is the work of making a particular assembly of dies actually function as a product.

5. Process Choice Becomes a Subsystem Decision

Chapter 1.6 noted that separate dies can use separate processes. Composition sharpens that from a possibility into a design axis.

When dies are built to be recombined, choosing where each one is manufactured stops being a consequence of the partition and becomes a decision made per subsystem, on that subsystem's own merits. A die dominated by dense logic and one dominated by interface circuitry no longer have to agree.

Keep the claim qualitative — what any specific function gains depends on the design, the process, and the product, and mapping functions to particular nodes is not something to assert casually. The structural point stands on its own: process technology becomes a subsystem-level architectural choice rather than one whole-system commitment.

6. The Interface Becomes a Published Contract

Here is the tension that produces everything after this module.

A private interface is perfectly adequate while one team designs both ends. They can change it whenever they like, document it loosely or not at all, and rely on shared assumptions — because the only parties who need to agree are already in the same meeting.

The more reusable a die is meant to be, the less that works. A die intended to meet partners it was not co-designed with needs its boundary specified rather than merely implemented: what the physical connection expects, how data is transported, how the link is brought up, how it is managed and monitored, how it behaves when something fails, and what counts as a compatible partner. Each of those is an assumption that stops being shared the moment the other end is designed by someone else — even a different team in the same company.

So reuse creates a requirement that decomposition never did. Which leaves the question this module has been building toward:

If chiplets are meant to be reusable building blocks, how do independently designed dies agree on what a connection means?

Module 3 answers that. This chapter only needs you to see why the question became unavoidable.

7. Common Misconceptions

Beliefs to unlearn before Module 2

"Every multi-die product is a chiplet architecture." No. Splitting a product into dies that exist only for that product is decomposition. Chiplet architecture requires the dies to be intended as reusable, composable building blocks — an intent visible mainly in whether the interface was specified for partners it was not co-designed with.
"A chiplet is just a physically small chip." Size is not the property. A small die drawn for exactly one neighbour is not a building block; a larger die with a stable interface reused across products is. What matters is functional coherence, an independently manufactured and tested identity, a defined boundary, and composable intent.
"Chiplets eliminate integration work." They relocate it. Redesigning and re-closing a die per product goes away; integrating a particular combination, designing its package, validating it, and managing compatibility as pieces evolve does not. Composition is a different integration problem, not the absence of one.
"Chiplets exist because dies reached the reticle limit." The weakest common answer. The reticle limit is one pressure and most designs never approach it. Module 1 assembled several: scaling benefits arriving separately, yield falling with area, economics punishing area twice, and — independently of all of those — a single die coupling decisions better made apart.
"Chiplet architecture means multi-vendor interoperability." Not yet, and not necessarily. Most chiplet-style products today compose dies designed within one organisation. Cross-vendor composition is the stated ambition and the reason standardisation efforts exist; treat it as a direction rather than a description.
"Reusable silicon means no re-verification." No. A reused die still has to be verified in each new context: with different partners, in different configurations, against a different package, and under the combined system's corner cases. Its internal verification is amortised; its integration verification is not.

8. Understanding Check

9. Module 1 in One Chain

This chapter closes the module, so here is the whole argument in one line:

Scaling benefits stopped arriving together → lithography caps a conventional die → yield falls as area grows → economics punish area twice and demand volume → one die couples decisions better made apart → partition the system across dies → design those dies as reusable building blocks.

Read it as accumulating pressure, not a single cause. Each link contributes something the others do not — which is exactly why answering "why do chiplets exist?" with any single one of them produces a weak answer. The honest version is that several independent pressures pointed the same direction, and composition turned out to relieve more of them at once than decomposition alone.

10. What Comes Next

Module 1 explained why the industry arrived here. Module 2 makes the vocabulary precise, starting with the term this chapter has deliberately used loosely.

2.1 — What Is a Chiplet? — a definition strict enough to reason with: what qualifies, what does not, and how chiplet, die, IP block, and multi-die system differ from each other.

Browse the full path on the UCIe tutorials index.