Smartphones, game consoles, cars, and appliances contain small black parts often called chips. Inside them, tiny electrical switches perform calculations, remember information, and communicate with other equipment.
The material and components that make these actions possible are called semiconductors. A semiconductor can be controlled so that electricity passes through it easily in one state and less easily in another. By arranging enormous numbers of tiny switches, a device can display pictures, run a game, or perform AI calculations.
It is becoming harder to make a high-performance chip as one large circuit. Instead, designers are increasingly dividing it into smaller semiconductor parts with different jobs and combining them inside one package. This article explains what becomes easier and which problems remain.
Why is one large piece difficult to make?
Semiconductor circuits are made by arranging many copies of a pattern on a round plate of silicon. The plate is then cut into separate rectangular pieces. That round plate is called a wafer.
Tiny defects or dust can enter during manufacturing. The larger a circuit becomes, the more likely it is that a problem will fall somewhere inside it. If one fault makes the entire large circuit unusable, fewer good parts are produced and the price can rise.
The machines that print circuit patterns onto the wafer also have a limit to the area they can draw at one time. As designers make circuits larger to gain performance, both design and manufacturing become more difficult.
One answer is to separate the parts that calculate, remember information, and communicate with the outside. Each is manufactured and tested, and parts that pass are placed together in one package. These smaller semiconductor parts are called chiplets.
Chiplets do not all have to come from the same newest factory process. Very fine circuits can be used where fast computation is needed, while a well-established process can be used for communication. Designers can choose a manufacturing method suited to each job. It may also become possible to update one part without rebuilding everything.
Small parts need rules for talking to one another
Splitting a chip does not make the pieces work as one computer. The parts need agreements. Where will a signal be sent? How fast? What happens when an error occurs? In what order do the parts start after the power is switched on?
If every company uses completely different rules, connecting parts made by different companies is difficult. Universal Chiplet Interconnect Express was created as a shared set of rules for connecting semiconductor parts. It is known by its initials, UCIe.
UCIe defines how chiplets send signals and data inside one product package. It covers more than electrical signaling. It also addresses the order in which data is carried and tests used to check whether a connection works correctly.
Even after shared connection rules are agreed, small semiconductors made by different companies cannot necessarily be freely combined right away. Other conditions still need to match, including physical size, electricity use, how heat escapes, and how quality is tested.
How did UCIe 3.0 change speed?
UCIe 3.0, the third major version of the specification, was released in August 2025. It added support for transfer rates of 48 GT/s and 64 GT/s.
GT/s measures how many signal transfers can be made in one second. At 64 GT/s, there are 64 billion transfers each second. This does not directly mean that 64 gigabytes of data can be sent each second. The useful amount of data changes with the way signals are recorded, the number of connections, and the extra information used to correct errors.
The highest transfer rate in the previous UCIe 2.0 specification was 32 GT/s. UCIe 3.0 doubled the maximum number of transfers.
For an AI chip, a faster calculation unit is not enough. If it waits for data to arrive from memory, work does not move forward during the wait. Many chiplets need a wider path that carries more information quickly between them.
Faster signaling can also use more electricity, create more heat, and make signals harder to keep clear. UCIe 3.0 includes mechanisms that adjust a connection while it is running and reduce unnecessary electricity use. It attempts to manage power rather than increasing speed alone.
Learning the condition of parts and reporting problems
When many chiplets are combined, a system needs to learn which parts are working correctly. It must be possible to check their condition and handle startup, testing, and problems. This is what making a system easier to manage means.
UCIe 3.0 defines a way to send a basic program to a part before the whole computer has fully started. This basic program is called firmware.
It also defines a way to give urgent management signals priority over other signals. There are methods for quickly reporting an emergency shutdown and for delivering time-sensitive signals with little delay.
The maximum distance for an auxiliary channel that carries management signals was extended to 100 millimeters. Ten centimeters is short in daily life, but it is long inside a small semiconductor package. The change gives designers more freedom in placing parts.
UCIe 3.0 is designed to understand the rules of earlier UCIe versions. This quality is called backward compatibility. It still does not guarantee that every part will connect. Designers must check that the speeds and optional features match.
What else must match besides the connection rules?
A real product must account for chiplet size. It must decide how electricity reaches each part and how heat escapes. A base and wiring are needed to hold and connect the parts.
Quality testing is essential. When something fails, engineers need to find which part caused the problem. If different companies are involved, they also need agreements about design secrets and responsibility for a failure. UCIe aligns an important part of the connection, but it does not solve all these questions at once.
Shared rules still matter. AI computation, memory, and communication may be combined without putting everything into one enormous circuit. Designers may find it easier to build a chip for a particular purpose and to replace one section with a newer generation.
The highest speed is not the only thing to watch. Do parts from different companies pass the same tests and operate reliably in real products? Does the design still offer an advantage after electricity for cooling is included? Does the range of parts that can be combined grow? UCIe 3.0 is not a finished marketplace of interchangeable parts. It is a shared starting point for many companies trying to build one.