Chiplet vs monolithic: what is the difference?
For decades the default was to integrate more functions onto one system-on-chip. As leading-edge wafers became more expensive and the largest designs approached the lithography reticle limit, many processor makers moved to chiplets. The choice now depends on die size, volume, the mix of functions and the packaging budget.
The short answer
A monolithic design puts every function on one piece of silicon, which gives the fastest and lowest-power on-chip communication but becomes expensive as the die grows, because yield falls with area. A chiplet design splits the system into several smaller dies, possibly made on different process nodes, and joins them in an advanced package. It improves yield and flexibility at the cost of die-to-die links and packaging complexity.
Side-by-side comparison
| Aspect | Chiplet design | Monolithic design |
|---|---|---|
| Concept | System split into several dies, joined in one package | All functions integrated on one die |
| Yield | Small dies each have higher yield; defects scrap less silicon | Yield falls steeply as die area grows |
| Process node | Each chiplet can use the node that suits it, e.g. leading edge for compute, older node for I/O | Whole die uses one process, including blocks that gain little from it |
| Size limit | Total silicon can exceed one reticle | Bounded by the reticle field of the lithography tool |
| Communication | Die-to-die links add latency, power and area for interface circuits | On-die wiring: lowest latency and energy |
| Design and verification | Interface standards, system partitioning, multi-die test and thermal analysis | Single-die flow, simpler to verify |
| Packaging | Needs advanced packaging such as 2.5D interposers, bridges or 3D stacking | Standard flip-chip or wire-bond package is usually enough |
| Reuse | Proven dies can be reused across products | New mask set for each design variant |
| Typical use | Server CPUs, large GPUs and AI accelerators | Microcontrollers, mobile SoCs, analogue and many mid-size chips |
Why yield favours chiplets
Random defects land on a wafer at a roughly fixed density. The chance that a die contains none of them drops faster than linearly as its area increases, so a very large die can have a poor yield even on a mature process. Splitting the same logic into several smaller dies means a defect discards only one small die, and each chiplet can be tested before assembly.
AMD's second-generation EPYC processors, launched in 2019, are a well-known example: compute dies made on a 7 nm process were combined with a separate I/O die on an older 14 nm process, so only the parts that benefited from the newest node used it.
What chiplets cost
Every signal that crosses from one chiplet to another needs interface circuits on both sides and travels farther than on-die wiring, adding latency and energy per bit. The package must provide enough wiring density for those links, which often means an interposer, embedded bridges or 3D stacking, each adding cost and assembly steps.
Design also gets harder. Architects must decide where to cut the system, verify behaviour across dies, manage power delivery and heat in the package, and ensure each die is a known good die before assembly.
Standard interfaces: UCIe
Early chiplet products used proprietary die-to-die links. Universal Chiplet Interconnect Express (UCIe), an open specification first released in 2022 and backed by a consortium of chip, foundry, packaging and cloud companies, defines a common physical layer and protocol stack for die-to-die connections in both standard and advanced packages. The aim is to let chiplets from different suppliers be combined, though multi-vendor ecosystems are still developing.
When monolithic still wins
For small and mid-size dies the yield gain from splitting is small, and the extra package cost and interface overhead outweigh it. Microcontrollers, analogue and power ICs, and most smartphone application processors remain monolithic. Latency-critical designs may also prefer one die. Chiplets pay off mainly in large, high-value devices where yield, reticle limits or mixed process nodes dominate.
FAQ
Is a chiplet the same as a multi-chip module?
The idea is related: both put several dies in one package. The term chiplet usually implies dies designed from the start as parts of one system, connected by high-density die-to-die links in an advanced package, rather than independent chips placed side by side.
Do chiplets always reduce cost?
No. They reduce silicon cost mainly for large dies on expensive nodes. For small dies the added packaging, interface area and test cost can make a chiplet design more expensive than a monolithic one.
Can chiplets be stacked vertically?
Yes. Chiplets can sit side by side on an interposer or bridge (2.5D) or be stacked on top of one another with micro-bumps or hybrid bonding (3D). Intel's Meteor Lake, for example, places several tiles on a base die using Foveros 3D packaging.
Terms compared
- ChipletA small die that implements part of a system and is combined with other dies in one advanced package, instead of building a single large monolithic chip.
- SoC (System on Chip)A single chip that integrates the main parts of a computer, such as CPU, GPU, AI unit, memory controller and modems; used in phones, cars and IoT.
Japanese companies involved
- Toray Engineering · Flip-chip bonders, die bonders and visual inspection systems
- Tokyo Electron (TEL) · Coat/develop, etch and deposition equipment
- DISCO · Dicing, grinding and polishing equipment
- IBIDEN · IC package substrates (FC-BGA)
- Socionext · Fabless custom SoC design
- Renesas Electronics · Microcontrollers, automotive SoCs, analog and power chips
More packaging comparisons
Sources
Semiconductor comparisons
- All semiconductor comparisons
- Hybrid bonding vs flip chip: what is the difference?
- TSV vs wire bond: what is the difference?
- Wafer test vs final test: what is the difference?
- HTOL vs burn-in: what is the difference?
- SEM vs TEM: what is the difference?
- XPS vs EDX: what is the difference?
- Czochralski vs float zone: what is the difference?
Last updated: November 2, 2026