²ÝÁñÉçÇø

So Many Dimensions, So Many Decisions

Because multi-die systems have so many interdependencies, every chip design choice should be made from a system-wide perspective to address the impacts that could permeate through the system. Power and performance analysis must be conducted at a system-level standpoint, given the new dimensions that these systems bring to the architecture design space. For instance, at 3D, thermal and power delivery problems tend to be more severe because it becomes much more difficult to dissipate heat. You¡¯d need to find a way to transfer the power through the lower level dies up to those at the top level to remove the power dissipation heat.

There are a host of considerations for multi-die systems, from the different ways to realize the interfaces to the choices of protocols, whether to place the dies next to one another or stack the dies vertically, what type of packaging to use, and much more. What are deemed the right choices will be driven by the power, performance, function, cost, and thermal considerations of the target application.

Key decisions fall into two main buckets. First, there are the multi-die system-architecture decisions. Here, designers could opt for aggregation, assembling the multi-die system from dies (or chiplets), or disaggregation, partitioning an application onto multiple dies. Design teams must also choose the protocols, location, and dimensioning of the die-to-die interfaces, the process technologies for each die, as well as the packaging technology. Then, there are the SoC-level architecture decisions. These choices involve hardware/software partitioning; IP selection, configuration, and connectivity; interconnect/memory dimensioning to configure the shared interconnect and memory subsystem; and system-level power analysis.

Design partitioning provides an illustration of the differences between monolithic SoCs and their multi-die system counterparts. In a monolithic SoC, designers must decide which functionality goes into which component, which subsystem handles the processing, and what type of IP to include to process the functionality. They also need to partition the data, determining where to store it and also how to connect all of the components so they can efficiently access the data they need.

A multi-die system adds many more dimensions to the design space, more degrees of freedom, if you will. Consider the cutlines between the different dies in the system¡ªif designed improperly, these can introduce bottlenecks, as die-to-die communication is afflicted with higher latency and reduced bandwidth compared to unconstrained on-chip communication. At this early stage in a design, however, it can be difficult to know the implications of all the decisions. The outcome of the first partitioning step is the top-level design spec. From then, it¡¯s on to implementation, which requires quite a lot of effort. The final performance of the system only becomes clear later downstream.

How Virtual Prototyping Tools Uncover Architecture Insights

One of the biggest challenges at the architecture planning phase is, so many important decisions must be made at the start of a project, when only a minimal amount of design data is available. It is also quite risky to rely solely on static spreadsheet analysis, given that many key performance indicators (KPIs), such as performance, power, and thermal, depend on dynamic effects of the application workload running on limited processing and communication resources.

How can design teams overcome this particularly vexing challenge?

Virtual prototyping of the multi-die system architecture presents an answer. Like an early electronics digital twin, a virtual prototype can be used to analyze the impact of architecture design decisions. The process begins with a definition of the workload and architecture: Which applications need to be executed? What are the non-functional requirements, like end-to-end latency and throughput? Which hardware components should execute this workload, and what are their power and performance properties?

For a realistic estimation of what the answers to these questions mean for the multi-die system, these functional and non-functional requirements need to be translated into the physical hardware properties of the system. The properties include the target process technology for the dies, their area, and the aspect ratios of the different components.

From this initial assessment, the design team can progress to a virtual assembly of the multi-die system. During this step, there are some important questions to answer: How are the components partitioned into dies?? How are the workload and data structures mapped onto these components? The answers together determine the traffic across the die boundaries. ?

There are, of course, many trade-offs to consider. For example, larger on-chip memories and caches are more expensive in terms of die area, but can help to reduce die-to-die traffic. ?By building an executable model of how the workload is running on the multi-die system architecture resources, the chip architects can collect a lot of data on how the actual system will behave. There are many key considerations here, including:

• What is the resulting utilization of resources?
• What are end-to-end latencies to execute a certain task?
• How much data needs to be transferred between the different chiplets?

Based on all the analysis, the team can modify all the aspects of the system architecture, quickly arriving at a feasible specification that meets the product requirements in an optimized way.