Chiplet

Imagine building a large piece of furniture as one single, giant moulded piece versus assembling it from several smaller, separately manufactured modules bolted together. If even one small flaw in the giant moulded piece ruins the whole thing, while a flaw in one module only means replacing that module, the modular approach starts to look considerably smarter as the overall piece gets more complex. Chiplets bring exactly this logic to chip design.

For decades, the default approach was building a single monolithic chip, one continuous piece of silicon containing every function the final product needed. As chips grew more complex, this became an increasingly expensive gamble against yield rate, a single defect anywhere on a very large, very complex monolithic die ruins the entire, expensive chip. A chiplet-based design instead splits that same overall functionality across several smaller dies, chiplets, manufactured separately and then connected together inside a single package using advanced packaging techniques.

The yield economics are the real payoff. Smaller dies are statistically far less likely to contain a fatal defect than one giant die covering the same total area, so a chiplet approach can achieve meaningfully better overall yield and lower cost than trying to manufacture the same functionality as a single monolithic chip, especially at leading-edge, defect-prone process nodes.

Chiplets also allow a manufacturer to mix and match, pairing a chiplet made on an expensive, cutting-edge process for the parts that genuinely need it with other chiplets made on a cheaper, more mature process for parts that do not, rather than paying leading-edge prices for the entire chip regardless of whether every part of it actually needs that level of performance. AMD's processor architecture has been one of the most visible commercial examples of this approach at scale.

Whenever a new high-performance chip is described as having multiple dies or a disaggregated architecture rather than being a single continuous piece of silicon, chiplets, and the yield economics driving their adoption, are almost always the underlying reason for that design choice.