Artificial Intelligence Can Help Optimize EDA Designs

AI-Based Analog Optimization

AI-based approaches do well in addressing optimization problems where traditional algorithmic approaches fall short. AI holds the potential to automate the manual loops in the design process. Much like a human designer, AI performs and learns from experiments, combining learnings across each of the experiments to understand and navigate in the solution space. In general terms, this approach is called sample-based optimization.

Sample-based approaches such as grid search, i.e., parameter sweeping, and random search, i.e., Monte Carlo simulation, have traditionally been used to aid designers during the analog design process. However, these approaches don’t scale well. The number of samples required for sufficient solution space coverage scales exponentially with design complexity.

More efficient general methods do exist, such as Bayesian Optimization, which is widely used in machine-learning applications. A Bayesian Optimizer builds a probability model of the objective function and uses it to select new sample points with high probability of scoring well in the metric space. As such, it takes learnings from previous samples to build a model that helps select future samples.

An AI-based approach represents an even more focused, intelligently directed way to navigate a large and complex solution space to find sample points that meet the specification. An AI capability can be devised as a sample-based optimization system that dynamically learns about the problem it’s tasked to solve​.

Such an AI approach can use actual, multi-corner/multi-testbench simulations to drive exploration of complex corner and testbench dependencies​. It can dynamically navigate process corners to reduce the number of simulations required, while converging across all corners. Through this process, the AI tool learns from its simulation experiments, using a live feedback loop to converge toward a solution that meets the specification.

A key advantage of such an AI system is that it doesn’t depend on any specific form of the problem it’s optimizing. However, unlike less efficient sample-based approaches, it will more efficiently self-adapt to the underlying objective. It also doesn’t optimize a proxy, but rather is driven by the actual circuit simulation.

Such a system is possible because the AI system makes informed decisions based on the experiments that it runs, reinforcing its internal perspective of the problem and objectives, which enables fast convergence.

AI-Based Analog Migration

Macro trends, including the slowing of Moore’s Law, manufacturing capacity constraints, and a challenging geo-political climate, are driving the need for newer capabilities to rapidly move designs between process nodes.

To take advantage of market opportunities and be resilient to supply-chain challenges, it’s essential for semiconductor companies to maneuver the supply-chain landscape with agility, including porting products from one foundry to another and from one technology node to another. While AI can help accelerate and automate circuit optimization in general, it holds a particular advantage during design migration.

As illustrated in Figure 1, the analog design migration process starts with a reference design, with specification, schematic, and layout in a given technology node, and ends with a completed and functional layout in the target node. The challenge of migrating an analog circuit from one node to another differs from general analog circuit design, in that the circuits are based on a prior version of the design.

This is good news for AI: Any design that’s been optimized in one context holds valuable learnings which are useful, even if the context, such as the technology node, has changed.

Figure 2 illustrates an AI-driven, automated design migration process. The first step is to migrate to the target node. Circuit elements and transistors are mapped to equivalent elements in the target node, the specification is adapted to the target, and the design is parametrized with parameters required to adjust the circuit to the specifications.