Chiplet design offers semiconductor teams a modular alternative to integrating every system function on one monolithic die. As advanced-process development becomes more demanding, placing selected functions on separate die can provide options for managing design scope, combining technologies and creating product variants. For industrial automation and other data-intensive systems, the approach also raises practical questions about interconnects, packaging, validation and system architecture.
Why consider chiplet design?
A monolithic system-on-chip (SoC) can offer compact integration, but it may be costly and time-consuming to develop and validate, particularly when a design is large or intended for several markets. A change to one part of a monolithic device can require redesign and re-verification of the full chip. Moving a design to a new process node can also require substantial work, even when many functions remain unchanged.
A multi-chip module (MCM) takes a different approach: it places two or more die in one package. Each chiplet can be designed, manufactured, tested and debugged separately before assembly. This divide-and-conquer method lets a team work on only the functions that need to change, while retaining other chiplets for later product versions where appropriate.
Chiplets can also be developed over time. A design team might create a new chiplet to add a feature or target another application without rebuilding every part of the system. The trade-off is that the module still needs a workable package, assembly and test strategy, and the economic case depends on the specific design.
Combining functions and semiconductor processes
An MCM can bring together identical chiplets, different types of logic, or a mix of digital and analog devices. Potential functions include processors, memory, FPGA resources, high-speed SerDes, and high-performance analog-to-digital or digital-to-analog converters. The die do not all have to use the same semiconductor material or fabrication process. For example, specialized chiplets can use materials such as silicon germanium, gallium arsenide, gallium nitride or indium phosphide where their properties suit a particular function.
Putting die in a common package can shorten connections compared with using separate packaged ICs linked across a printed circuit board. Those shorter connections can support higher bandwidth and lower latency, while avoiding some of the power associated with board-level communication. The package-level connection is not automatically the right answer for every design: the required bandwidth, power, latency and manufacturing cost all need to be considered together.
Chiplet interconnect remains a system-level decision
For chiplet-based systems to be repeatable across suppliers and designs, the interconnect must be more than a set of one-off connections. A die-to-die interface needs to meet the application’s bandwidth, latency and power requirements, while also fitting the physical package and substrate. Designers must also consider how to identify and handle known-good die before assembly.
Possible physical approaches include parallel signaling, multiplexed pins and SerDes links. Existing protocols such as PCIe and Ethernet are familiar options, but their additional protocol functions can bring latency, power and die-area overhead that may not suit every chiplet connection. Lighter-weight approaches may reduce overhead, but need suitable error handling and broad adoption before they can serve as common interfaces.
The broader challenge is coordination across the protocol stack: physical-layer signaling, link behavior, memory management and application-level communication must work together. The 2019 discussion of chiplets described industry efforts to develop open specifications that could combine existing standards with new interfaces where gaps remained. The practical lesson for system architects is to assess interoperability and ecosystem support alongside raw interface performance.
Where FPGA chiplets fit
An FPGA chiplet can add programmable logic to a package without requiring that logic to be embedded in the main ASIC or SoC. This approach is distinct from integrating eFPGA IP directly into a monolithic design, and the two options are not mutually exclusive. The choice depends on the desired integration, implementation constraints and how the programmable function will be used.
One potential benefit of package-level integration is reducing the distance between the FPGA and neighboring devices. Compared with communication between separately packaged chips over a board, a chiplet connection can provide tighter integration and reduce interface power. The source whitepaper gives an illustrative comparison for a 1 Tbps connection: approximately 10 pJ per bit over PCB traces between packaged chips versus approximately 1 pJ per transition for a similar MCM connection. These figures describe that example, not a universal result for all designs.
Choosing a flow-through or sidecar architecture
Flow-through: process data in the datapath
In a flow-through architecture, high-speed data passes through the FPGA as it is processed. This can suit data-plane tasks that require rapid handling, protocol conversion or interface adaptation. Processing data along its path can reduce movement between memory and compute resources. If the FPGA connects directly to package pins, it needs appropriate I/O capabilities or support from other chiplets that provide the required external interfaces.
For example, an FPGA chiplet could handle programmable protocol or exception processing in a network interface, while a separate NIC chiplet manages other functions. A processor can supervise the devices without directly handling the high-speed traffic. Similar partitioning can combine FPGA logic with converters in signal-processing systems.
Sidecar: accelerate a neighboring device
In a sidecar architecture, the FPGA communicates with one other chiplet, often a processor, to accelerate selected workloads. Because it does not need to serve as a path for external I/O, the FPGA may not require the same large I/O structures as a flow-through implementation. A sidecar can also be designed as an optional module feature: a product family may offer versions with or without the accelerator, with corresponding differences in chiplet, assembly and test costs. Hybrid designs can combine aspects of both arrangements.
Practical implications for design teams
Chiplet design is most useful when modularity solves a specific engineering or product-planning problem. Before committing to an MCM, teams should evaluate:
- Which functions benefit from separate die, including any differences in process or semiconductor material.
- Whether the workload calls for flow-through data processing, sidecar acceleration or a hybrid.
- The bandwidth, latency, power and error-handling requirements of each die-to-die link.
- Package, substrate, assembly, testing and known-good-die requirements.
- Whether the chosen interfaces and chiplets can support planned product variations.
Chiplets do not remove the need for careful system integration; they change where that work happens. For automation and other demanding applications, the practical opportunity is to partition a system around its functions and interfaces rather than treating a single large die as the only design target. When packaging, interconnect and validation requirements are addressed early, an MCM can offer a flexible route to combining compute, programmable logic, memory and specialized I/O.