Chiplets and UCIe Explained: How Modern Processors Are Built
What a chiplet is, why chipmakers abandoned monolithic dies, and how the UCIe standard (versions 1.0 to 3.0) connects chiplets, with PCIe compared.
A chiplet is a small piece of silicon that does one defined job, such as CPU cores, I/O or memory control, and is combined with other chiplets inside a single package to work as one processor. UCIe (Universal Chiplet Interconnect Express) is the open standard that defines how those chiplets talk to each other, so dies from different companies can share a package. The latest UCIe specification with published details on the consortium's site, as checked on 30 October 2026, is version 3.0, released in August 2025.
What is a chiplet?
Traditionally a processor was one monolithic die: every block, from cores to memory controllers, was etched on the same piece of silicon with the same manufacturing process. A chiplet design splits that die into several smaller ones and connects them in the package. The chiplet overview on Wikipedia describes each chiplet as a well-defined subset of functionality, assembled with others on an interposer or substrate into a complex part.
The result still looks like one chip to the buyer and to software. Inside, it is closer to a small board built at the scale of micrometres.
Why chipmakers moved away from monolithic dies
Three pressures push designs toward chiplets.
- Reticle limit. A lithography scanner can expose only a limited field in one shot, roughly 26 by 33 mm (about 858 mm2) for current EUV tools. No single die can be larger than that. Nvidia describes Blackwell as two reticle-limit dies joined into one GPU, which shows how the limit shapes the biggest chips. Newer high-NA tools expose an even smaller field, which we cover in our High-NA EUV explainer.
- Yield. A defect anywhere on a large die can kill the whole die. Small dies can be tested separately and only good ones packaged, a practice known as known good die.
- Cost and mixing of nodes. Logic benefits from the newest and most expensive process, while I/O and analog circuits gain little from it. A chiplet package can use an advanced node for compute and an older node for the rest. Our guide to TSMC process nodes shows how large the price gap between nodes can be.
The trade-off is power and complexity. Dies packed close together concentrate heat, and every link between chiplets costs some latency and energy compared with wires on one die.
Chiplet architecture in real products
- AMD. Ryzen desktop processors based on Zen 2 and later use separate core chiplets plus an I/O die, according to the Wikipedia overview. AMD introduced the approach to desktop with 3rd Gen Ryzen in 2019, and EPYC server chips use the same idea at larger scale.
- Intel. Meteor Lake and later Core Ultra chips are built from tiles, which are Intel's term for chiplets, stacked or placed on a base using its Foveros packaging technology.
- Nvidia. Blackwell joins two large dies with a 10 TB/s link, and the chip has 208 billion transistors, per Nvidia's announcement.
- Apple. M1 Ultra joins two M1 Max dies with UltraFusion, which Apple says provides 2.5 TB/s of inter-processor bandwidth.
These links are proprietary. A chiplet from one vendor cannot be dropped next to another vendor's die. That is the gap UCIe targets.
How chiplets are connected: packaging options
The package decides how many wires can run between dies, and therefore how fast and efficient the connection is.
- Organic substrate (2D). Dies sit side by side on a substrate with coarser wiring. It is the cheapest option and is what AMD's Ryzen desktop parts use.
- Silicon bridge. A small piece of silicon embedded in the substrate provides dense wiring only where two dies meet.
- Silicon interposer (2.5D). Dies sit on a larger silicon layer with fine wiring. TSMC's CoWoS family is the best-known example; see What is CoWoS? for the differences between its variants. It is also how GPUs connect to stacked memory, covered in our HBM4 vs HBM3E piece.
- 3D stacking. Dies are placed on top of each other. Hybrid bonding joins them copper to copper with very small bump pitches, which gives the densest connections.
What is UCIe?
UCIe is an open die-to-die interconnect standard governed by the UCIe Consortium. Its aim is an ecosystem in which a buyer can combine chiplets from different suppliers in one package. The consortium was announced in March 2022. Per Wikipedia, the promoter members are AMD, Arm, ASE Group, Google Cloud, Intel, Meta, Microsoft, Qualcomm, Samsung and TSMC, and Alibaba and Nvidia joined the board in August 2022.
The specification covers the physical layer, a protocol stack, a software model and compliance testing. Its protocol layer builds on PCIe and CXL, so existing software can treat a chiplet link much like a familiar interconnect.
UCIe specification versions
Release dates below come from Wikipedia's UCIe article; feature lists come from the consortium's specifications page.
- UCIe 1.0 (March 2022). Defined the die-to-die layers and support for advanced packaging, with the aim of building systems larger than the reticle limit.
- UCIe 1.1 (August 2023). Backward compatible. Added runtime health monitoring and repair, support for several protocols at once, and bump maps for lower-cost packaging.
- UCIe 2.0 (August 2024). Added optional manageability, a test and debug architecture, and UCIe-3D for hybrid-bonded stacks.
- UCIe 3.0 (August 2025). Added 48 GT/s and 64 GT/s data rates, double the 32 GT/s of version 2.0, plus a longer sideband channel of up to 100 mm and features such as early firmware download.
Every version is backward compatible with the earlier ones. The consortium's specifications page also lists a UCIe 4.0 entry, but when checked on 30 October 2026 it carried no feature details.
UCIe vs PCIe
The two are related but work at different scales. PCIe connects separate components across a motherboard or through a slot, over distances measured in centimetres. UCIe connects dies inside one package, over millimetres, where much denser and more energy-efficient links are possible. UCIe reuses PCIe and CXL protocols on top of its own physical layer rather than replacing them. You can see how fast PCIe itself has grown in our PCIe generations guide.
What UCIe changes, and what it does not
A shared standard lowers the cost of reusing a chiplet and lets smaller companies sell specialised dies, such as an accelerator or an I/O block, to be packaged next to a larger vendor's compute die. It also makes testing and security more consistent through the manageability and trust work added since version 2.0.
It does not remove the hard parts. Packaging capacity, thermal limits and die-to-die testing remain bottlenecks, and the biggest chips from AMD, Intel, Nvidia and Apple still use their own tuned interconnects. Expect UCIe to appear first in custom silicon and in links between dies from a single company, with multi-vendor packages growing more slowly.
For more on the hardware side, see our Chips coverage and the wider Industry section.
Frequently asked questions
What is a chiplet in simple terms?
A chiplet is a small die that performs one function, such as compute or I/O, and is packaged together with other chiplets to act as a single processor. It replaces one large die with several smaller ones.
What does UCIe stand for?
UCIe stands for Universal Chiplet Interconnect Express. It is an open standard for connecting chiplets inside a package, maintained by the UCIe Consortium.
Is UCIe the same as PCIe?
No. PCIe links separate devices across a board, while UCIe links dies within one package. UCIe's protocol layer is based on PCIe and CXL, which is why the names are similar.
What is the latest UCIe specification?
UCIe 3.0, released in August 2025, was the latest version with published details when this page was checked on 30 October 2026. It supports data rates up to 64 GT/s. Check the consortium's specifications page for newer releases.
Why do chipmakers use chiplets?
Chiplets get around the reticle size limit, improve yield because small dies can be tested before assembly, and let designers mix manufacturing nodes in one package.