Chiplet

A chiplet is a design approach in which, instead of making one large chip, you make small chip pieces (dies) separately by function and then join them as one inside a package. Each piece can be produced on the process that suits it best, which makes the design more flexible.

This approach drew attention as fine semiconductor processes got harder and the area of a single chip hit limits. It came into wide use as the need grew to protect yield and lower cost when making large AI accelerators.

In a chiplet design, dies with different roles, such as compute, input and output, and memory, are placed side by side on a silicon interposer or an organic substrate. The dies are joined by short wiring and a set communication standard, so that they work like one large chip inside the package.

When dies exchange data, latency and power use matter most. So if a standard such as UCIe matches the signaling method and the transfer speed, dies made by different companies can be attached together.

It is widely used for large chips, such as high-performance AI accelerators and server processors. If you print one big chip, a defect often forces you to throw it away, but if you split it into small dies, you discard only the bad part and protect yield.

To use it, you must design the connection standard between dies, the packaging method and heat management together. When you mix dies from several suppliers, specification compatibility has to be confirmed, and the cost of the substrate and package must be weighed too.

It is easy to assume that chiplets always improve performance and cost. Communication between dies is slower and uses more power than wiring inside a single chip, so with many connection hops you can end up worse off.

The packaging process is complex, so the defect rate and cost at the assembly stage can also rise. Chiplets change how you design and build, so there is no reason to adopt them when a single chip is enough.

It is like making a big lunch box not all at once, but making each side dish separately and packing them into one box. Each dish is made in the kitchen that does it best, and the box serves them as one meal.

The larger an AI chip gets, the faster the cost and the risk of defects rise when you make it as a single chip. Knowing how chiplets work helps you see why news about the latest AI accelerators and high-performance chips focuses on die connections and packaging.