Chiplets are usually described as a manufacturing story: as shrinking a whole chip becomes harder and more expensive, designers break the system into smaller dies and assemble them in an advanced package. The framing that follows is a race for capacity โ€” who can build enough silicon interposers, who can secure substrate supply, who can book the packaging lines. That race is real, but it is not where the hardest problems now sit. The more consequential challenge is verification. Dies designed by different teams, and increasingly by different suppliers, must interoperate, be tested before assembly, share power and thermal budgets, and carry evidence of where they came from and what they are. Those are engineering requirements, not back-end assembly details, and they will decide which chiplet designs are manufacturable at scale.

For the Global Tech Awards, which recognises engineering across the technology sector, the shift sits where semiconductor design, test engineering, supply-chain management and security meet. It rewards organisations that treat packaging as part of the design brief rather than the last step before shipment.

The unit of design has changed

In a monolithic chip, design and verification share a boundary. One team, or a tightly coupled group, owns the whole die; interfaces are internal; a fault found in test can be traced through a known design hierarchy. The chiplet model breaks that assumption. A system may combine a compute die from one supplier, memory controllers from another, an input-output die on a different process node and custom accelerators designed in-house. Each is a separate product with its own specification, test programme and owner.

It is no longer enough for each die to work in isolation. The assembled system has to work, and the failure may not belong to any single die. A marginal timing violation at an interface, a mismatch in power-state sequencing or a difference in how two suppliers read a specification can produce a fault that appears only once the dies are joined โ€” and the package is difficult and expensive to rework.

The practical consequence is that interface compliance becomes a product-level requirement. Teams need agreed electrical and protocol specifications, conformance tests both sides run, and a shared understanding of corner cases before tape-out rather than after assembly. Where standards-based die-to-die interfaces are used, conformance to the standard is necessary but not sufficient; the integration still has to be verified for the specific combination of parts.

Known-good-die testing moves to the centre

Assembly economics depend on yield. If a package contains several dies and one is defective, the whole package may be lost, so the cost of a bad die multiplies with the number of dies in the stack. That is why known-good-die testing โ€” establishing with high confidence that each die is functional before assembly โ€” becomes a commercial as well as a technical priority.

Doing it well is harder than it sounds. Bare dies are difficult to probe, particularly when they carry fine-pitch connections intended for the package rather than a test head. Some defects appear only at speed, under particular voltage or temperature conditions, or in combination with neighbouring circuitry. Coverage that was adequate for a standalone product may leave gaps once the die becomes one component of a larger system.

There is also an information problem. The organisation assembling the package often does not own the test programme for every die inside it. It depends on supplier test data, agreed escape rates and a shared definition of what counts as known-good. Procurement teams therefore negotiate test coverage, data formats and reporting obligations alongside price and lead time. A supplier that cannot provide credible test evidence is a different kind of risk from one that costs more.

Package-level coverage has to be planned at the same time. Some failure modes โ€” interconnect defects, thermal-interface problems, mechanical stress โ€” exist only after assembly, and die-level tests cannot reach them. The test strategy is therefore layered: die-level, interface-level and package-level, each with its own coverage goals.

Thermal, mechanical and power interactions must be co-designed

Advanced packaging is sometimes treated as a capacity constraint: enough lines, enough substrates, enough yield. Capacity matters, but the harder engineering is the interaction between dies that now share a physical environment. Heat generated in one die raises the temperature of its neighbours, and thermal gradients can change timing behaviour and reliability in ways no single-die model predicts. Power delivery has to reach multiple dies through a shared network whose impedance depends on the arrangement of the whole package.

Mechanical behaviour adds another layer. Materials with different coefficients of thermal expansion are bonded together, and the resulting stress affects warpage, joint reliability and long-term durability. These determine whether a package survives the temperature cycles of real deployment.

The design implication is that thermal and power budgets cannot be allocated after the dies are designed. They have to be negotiated while the architecture is still fluid, with each die's power envelope, operating range and duty cycle treated as inputs to a shared model. Teams that defer this work to the assembly stage tend to discover that the package, not the silicon, is the limiting factor.

Security and provenance travel with the package

Each die in a package may come from a different supplier, and each may carry firmware, keys or configuration data. For AI accelerators and infrastructure chips in particular, what is actually inside the package becomes a security question. Buyers want to know that components are genuine, that their provenance is documented and that the assembled system can attest to its own configuration.

That pushes lifecycle tracking and supply-chain evidence into the design. Identifiers that survive assembly, records that link a package to its constituent dies and mechanisms for verifying that a component has not been substituted all become part of the product definition โ€” and part of procurement, because a supplier that cannot provide traceable evidence is harder to qualify where trust in the bill of materials matters. None of this is exotic in principle; the difficulty is that it has to be designed in early, when identifiers, test access and data formats are being decided, rather than retrofitted once the package is in production.

What engineering and procurement teams should treat as requirements

The chiplet transition rewards organisations that move verification earlier and treat it as a first-class deliverable. In practice that means a few concrete shifts.

  • Interface agreements before design freeze. Electrical, protocol and timing specifications should be agreed with counterparties, with conformance tests both sides can run.
  • Test data as a supplier deliverable. Known-good-die evidence, escape rates and data formats belong in the contract, not in a later conversation.
  • Shared thermal and power models. Budgets should be allocated across dies while the architecture can still change.
  • Package-level test coverage planned up front. Interconnect, thermal and mechanical failure modes need their own coverage goals.
  • Traceability designed into the product. Identifiers and provenance records should be specified alongside function.

There is an opportunity here as well as a burden. Standardised interfaces and shared test methods can lower the barrier for smaller designers who cannot build every die themselves, widening the pool of organisations able to assemble competitive systems. But that benefit only materialises if interoperability and traceability are engineered rather than assumed. The chiplet era will be judged less on how many dies fit in a package than on whether anyone can prove, before and after assembly, that the system works and can be trusted.

โ† PreviousBack to Articles