Introduction
The design of integrated circuit (IC) cell layouts is a complex, multi-layered task that traditionally required significant manual effort from expert designers. In the early 1980s, as the complexity of Very Large Scale Integration (VLSI) grew, the need for automated tools became critical. TALIB was developed as a pioneering knowledge-based system to bridge the gap between high-level circuit schematics and physical mask geometry. By employing hierarchical planning and high-level subcircuit abstractions, TALIB automates the NMOS cell layout process, trading a modest increase in chip area for a substantial reduction in design time and human error.
Methodology
TALIB is implemented as a production system using the OPS5 AI programming language. Its architecture is built upon three primary pillars: hierarchical planning, a vast rule-based knowledge organization, and high-level abstractions.
Hierarchical Planning and Knowledge Organization
The system avoids the pitfall of manipulating thousands of individual geometric rectangles simultaneously. Instead, it partitions the layout surface into zones and the circuit netlist into topological groups. This "opportunistic" refinement allows the system to move from abstract subcircuits down to specific geometric primitives. The knowledge base consists of approximately 1,200 rules, categorized into subtasks (handling specific design steps like placement and compaction) and "demons" (global rules that monitor for constraint violations or propagate requirements).
Circuit-Layout Abstractions
A key innovation in TALIB is the use of "circuit-layout models" for common subcircuits, such as inverters or gates. Rather than treating a circuit as a flat sea of transistors, TALIB recognizes these functional clusters. These models contain essential attributes like estimated dimensions and terminal positions, enabling the system to reason about spatial constraints before finalizing the geometry.
For intricate local routing where heuristics might be insufficient, the system utilizes a modified Lee expansion algorithm to ensure connectivity while adhering to fabrication rules.
Evaluation and Key Results
TALIB was evaluated using a suite of twelve NMOS digital circuits, including a 36-transistor multiplier slice. The system successfully produced "correct-by-construction" layouts that satisfied all boundary and fabrication constraints.
The primary findings from these evaluations include:
- Efficiency: A valid layout for a 36-transistor circuit was produced in fewer than 2,000 rule firings.
- Quality: While the generated layouts are 10% to 35% larger in area than those crafted by human experts, they are produced significantly faster and are guaranteed to be error-free regarding design rules.
- Feasibility: The system proved that a rule-based AI could effectively manage the topological nuances of IC design by operating on functional clusters rather than raw primitives.
Conclusion and Future Work
TALIB demonstrates the power of applying AI planning techniques to physical design. While the area overhead is a notable drawback, the gain in design speed and reliability is a significant advantage in the context of large-scale chip design. Future work aims to reduce the area penalty by refining layout rules and adapting the system to modern technologies like CMOS. TALIB was envisioned as a component of a larger hierarchical design automation framework (CMUDA) capable of generating entire chips from behavioral descriptions.
Summary updated