Abstract
As semiconductor manufacturing approaches the physical and economic limits of monolithic System-on-Chip (SoC) designs, chiplet-based architectures have emerged as a viable alternative. However, the transition to chiplets introduces complex trade-offs involving Die-to-Die (D2D) overhead, advanced packaging costs, and increased design complexity. This paper introduces Chiplet Actuary, a comprehensive quantitative cost model designed to evaluate the economic feasibility of multi-chiplet architectures. By accounting for both Recurring Engineering (RE) and Non-Recurring Engineering (NRE) costs, the model identifies critical thresholds for volume, granularity, and technology reuse that determine the profitability of chiplet designs.
Introduction
The semiconductor industry is currently facing a dual challenge: the "reticle limit," which restricts the maximum size of a single die, and declining yields for large chips on advanced process nodes. While partitioning a monolithic SoC into smaller chiplets can improve yields and bypass size constraints, it is not a "free lunch." Chiplets impose a "tax" in the form of additional silicon area for D2D interfaces, sophisticated packaging requirements (e.g., 2.5D interposers), and higher initial design costs.
Architects have historically lacked a rigorous framework to quantify these trade-offs during the early stages of design. Chiplet Actuary fills this gap, providing a mathematical foundation to determine when and how to transition from monolithic to multi-chiplet architectures.
The Chiplet Actuary Cost Model
The model categorizes expenses into two primary streams to provide a holistic view of the product lifecycle:
1. Recurring Engineering (RE) Cost
RE costs represent the manufacturing expense per unit. Chiplet Actuary calculates this by factoring in:
- Silicon Area and Yield: Using Seed’s yield model to account for die defects.
- Packaging Complexity: Costs associated with Multi-Chip Modules (MCM), Integrated Fan-Out (InFO), and 2.5D Silicon Interposers.
- Assembly Loss: The cost of "wasted" known-good dies lost during faulty assembly processes.
2. Non-Recurring Engineering (NRE) Cost
NRE costs are the one-time investments required to bring a design to market. A key contribution of this model is the explicit inclusion of NRE for D2D interfaces, alongside traditional costs like IP licensing, mask sets, and verification. These costs are critical for determining the amortization threshold across production volumes.
Methodology and Validation
The model was developed and validated using a combination of industry-standard commercial databases (e.g., IC Knowledge LLC), public technical specifications from AMD’s Zen and EPYC architectures, and proprietary data from TSMC’s 7nm and 5nm process nodes. This multi-faceted data approach ensures that the model reflects real-world manufacturing constraints and pricing.
Key Architectural Insights
The Volume and NRE Determinant
One of the most significant findings is that for low-volume products (typically under 500,000 units), monolithic SoCs remain more cost-effective. The high NRE cost of designing multiple chiplets and their respective interfaces requires high production volumes to amortize. For a typical 5nm system, the "break-even" point where chiplets become cheaper than monolithic designs is approximately 2 million units.
Granularity and Diminishing Returns
There is a limit to the benefits of partitioning. While moving from a monolithic design to three chiplets can yield significant savings, further partitioning (e.g., to five or more chiplets) often results in diminishing returns. The marginal yield gains are frequently offset by the increased silicon area required for D2D overhead and the complexity of the interconnects.
The Power of Reuse
The most substantial cost savings are realized through Single Chiplet Multiple Systems (SCMS) strategies. By designing a single chiplet type that can be reused across an entire product stack—a strategy successfully employed by AMD—designers can drastically reduce NRE costs and accelerate time-to-market.
Results and Case Studies
Validation against AMD’s public data confirms that their chiplet strategy can save up to 50% in die costs. However, the model also highlights that packaging can account for up to 30% of the total cost in smaller configurations, such as 16-core systems. Additionally, the model demonstrates the benefits of heterogeneous integration: by using older, cheaper nodes (e.g., 14nm) for components that do not scale well, such as I/O, while keeping core logic on 7nm, designers can achieve a total cost reduction of over 10%.
Limitations and Future Directions
While Chiplet Actuary provides a robust framework for cost analysis, it currently focuses exclusively on economic metrics. Future iterations of the model will aim to integrate:
- Performance and Power: Accounting for the latency and power consumption of D2D interfaces.
- Thermal Management: Modeling the heat dissipation challenges of densely packed dies.
- Testing Protocols: Incorporating the specific costs of Known Good Die (KGD) testing.
By evolving into a holistic "PPA-C" (Power, Performance, Area, and Cost) tool, Chiplet Actuary will continue to guide architects through the increasingly complex landscape of modern semiconductor design.
Summary updated