What is a Multi-Die Chip Design?

Better Design Flexibility…and Economics

Traditionally, chipmakers have moved to smaller process nodes to achieve their desired power/performance goals, functionality, form factor, and cost. With an increasing need for processing power, however, SoCs are becoming quite large—beyond what can be fabricated with reasonable yield. We’ve also reached a stage where, in some cases, moving to an advanced node no longer makes sense. As die sizes approach the reticle limits of manufacturing equipment, it simply becomes uneconomical to physically accommodate all the logic, IO, and memory needed for compute-intensive applications. Instead, chip designers are splitting their designs into multiple smaller dies, which are easier to fabricate and produce better yields.

In short, a multi-die design is one where a large design is partitioned into multiple smaller dies—often referred to as chiplets or tiles—and integrated in a single package to achieve the expected power and form factor goals. While monolithic designs boast all of their functionality on a single piece of silicon, a multi-die approach provides the product modularity and flexibility to mix and match separate dies into packages to address different market segments or needs. For example, an end product that’s available in different markets targeting low- and high-end as well as mid-range segments can benefit from a multi-die approach. There’s also flexibility in terms of mixing process nodes in a multi-die design. For example, a processor for a compute function might be on an advanced node, while the IO function is on an older node, with both utilizing the technology node optimally.

The architecture for a multi-die design can come in different formats. Dies can be placed side-by-side and be connected by dedicated die-to-die interfaces, a prevalent and lower cost approach. For greater density, the blocks can be assembled in a 2.5D or 3D package. 2.5D designs integrating GPU and high-bandwidth memory (HBM), which feature 4 to 12 large HBMs on an interposer, have been a workhorse for artificial intelligence for a decade. They’re now finding ways to serve new end markets such as 5G infrastructure, data centers, and large networking systems.

Taking a System-Wide Perspective

The individual functions of a multi-die design are not so different from those of monolithic chips, with a few key exceptions:

• The connections between the dies are critical: they must be power efficient, have low latency, provide high bandwidth to transfer massive amounts of data between dies, and deliver error-free operation
• The system analysis, whether 2D or advanced 2.5D or 3D, will have to take into account coupling effects of through-silicon vias (TSVs), through-dielectric vias (TDVs), redistribution layers (RDLs), interposers, and substrates

Components aside, there’s also a need to approach multi-die designs with a much different perspective than is needed for monolithic chips. When an SoC is split up into separate modules, designers must take more of a system-wide view of performance and cost. Because multi-die designs add complexity to their monolithic counterparts, it’s important to co-design them with an upfront understanding of their thermal footprint, signal and power integrity, mechanical issues, routing considerations, and other key parameters.

Different packaging technologies enable different routing density for the inter-die connections, with different electrical characteristics. The die-to-die interface architecture is optimized to support the application performance targets with the selected packaging technology. For example, high-speed SerDes architectures are more aligned with the characteristics of 2D and 2.1D packaging, whereas high-bandwidth parallel architectures can better leverage the routing density offered by 2.5D and 3D packaging.

It’s important to co-design multi-die designs with an upfront understanding of their thermal footprint, signal and power integrity, mechanical issues, and routing considerations.

Let’s consider the example of TSVs that are used to create connections between the dies. There are new effects around TSVs related to noise, thermal cross talk, and spacing rules, for instance, that aren’t present in monolithic SoCs and must be considered relative to their impacts on the design as a whole. Doing so early on can help prevent costly yield issues later.

With signals being routed between the dies and within the packaging, die-to-die interface IP and high-speed serial links can support the process as dies that are split up are reintegrated into a new package. Speaking of die-to-die interface IP, the chip/package co-design process should also integrate these building blocks into the mix. For optimal power, thermal, and mechanical integrity, it’ll be important for disciplines to shift up in the design process.

For a better understanding of system performance and the system architecture, designers need a solution that can guide them on how to best bring different aspects of their design together in a holistic way that is optimized for power and signal integrity, thermal considerations, and noise.