Efforts to speed up public domain SPICE simulators developed at the University of California, Berkeley half a century ago provided the foundation for what would eventually become Cadence Design Systems. Today, the company still holds a dominant position in mixed-signal layouts.
Automation is largely dominated by digital design. Companies are now touting artificial intelligence (AI) agents as the next step. The slow progress in analog is not because we don’t want to try. A quarter of a century ago, companies like Barcelona Design Automation promised designers a way to automatically generate analog circuits from a set of specifications. But apart from porting circuits from one process node to another, analog designs have proven persistently resistant to automation, though not necessarily for obvious reasons.
Ensilica CTO Alan Wong sees the latest generation of generative AI tools coming out of academia and thinks there’s room for adoption. “The work was very good, so I’m thinking positively,” Wong said. “But integrating into the commercial world is the difficult part.”
Machine learning is making a difference so far with the same type of tools that created one of the two companies that merged in the late 1980s to form Cadence. It’s about accelerating the characterization and verification of manually created circuits and layouts. Machine learning efforts primarily rely on the use of neural networks as general function approximators rather than generative engines. The large number of inputs and internal connections within a neural network allows it to learn and emulate complex interaction behavior.
Solido Design Automation made its foray into machine learning less than a decade ago. The company focused on issues such as the inherent variability of transistors and estimating the effects of parasitic elements in the library under varying temperatures and voltages required for process design kits (PDKs). This work has sped up the effort to build the libraries needed to characterize circuits across the range of conditions that the hardware is expected to handle.
Since Siemens acquired Solido in late 2017, efforts have expanded to reduce the number of time-consuming simulations required to analyze the effects of process, voltage, and temperature (PVT) variations. Even with the speed-optimized FastSPICE tool, it takes time. A trained model can, in principle, avoid the need to run a large number of models.
As Amit Gupta, former CEO of Solido and now senior vice president of AI strategy at Siemens EDA, explained at the Design Automation Conference last summer, “We talk about how the cheapest simulations are the ones you don’t have to run. Rather than brute force, can you first reduce the number of simulations? That’s why we use reinforcement learning and machine learning.”
Expansion of adoption
It took time for designers to be able to use machine learning results to reduce simulations. However, the adoption of sensitivity analysis approaches is growing.
“In the past, we’ve run into situations where the model in the library model is missing a corner that we need. Using AI with the data points we have, we can predict new corner cases. The question is whether we can trust it enough to tape out,” Wong says. Therefore, detailed simulation remains a key element of the design pipeline.
An example of the potential savings from corner and PVT prediction is Neuron IP’s work on driver circuits for chiplet-to-chiplet communication. According to Prabhnoor Kainth, System Verification Engineer at Neuron, the Solido DE tool has reduced the number of randomized simulations required by IP developers. These simulations evaluate the effects of multiple process variations. This process typically requires thousands of Monte Carlo simulations with randomly selected combinations of parameters. Modeling using machine learning reduced the number of actual simulations from 30,000 to less than 1,000. The model built for the IP block allowed the design team to focus on the PVT variations that were most likely to affect the circuit’s operation. These can be performed using full simulation.
Generative AI is also being introduced into analog engineering, but not to the level previously seen in digital design.
British porting specialist Thalia Design Automation has incorporated the underlying technology of Transformers to easily transfer existing layouts from one process to a new target. The design’s Pretrained Transformer (DPT) extracts device and other attributes from the PDK provided by the foundry. That information feeds the company’s porting and layout automation tools with the parameters needed to determine how to optimize the shape and position of the device in the new process.
Language model token bulk text generation is fairly well suited for digital design and document extraction, as it can generate Verilog code. The more graphical nature of analog design contradicts that. Various groups have used graph neural networks to represent connections between devices, but these tend to overemphasize local links.
One of the benefits of Transformer is that it allows the effects of distant tokens to mirror each other. For AnalogGenie, a language model trained to find valuable circuit topologies, a team led by George Washington University assistant professor Weidong Cao translated the schematics into text. To do this, they employed an Eulerian circuit. These describe the graph as a list of nodes. Each is accessed only once through a direct connection. This structure allows for graph reconstruction. Another experiment, Stanford University’s AnalogCoder, uses a Python interface to SPICE as a way to process analog circuits in text.
The second problem is the lack of data. In the silicon design industry, very few people want to share their results. Ideally, they can share both good circuits and mistakes to further advance the model. Cao said at a seminar on efficient AI hosted by Rutgers University last fall that he had to rely on textbooks and published academic papers. This limits the depth of exploration that the generative AI can perform. Actual implementations may rely on the ability to fine-tune locally executed models using internal data.
Other options
Some radio frequency (RF) designs offer more options, including the ability to use more graphical AI techniques. The matching networks required for high frequency RF circuits use metal traces on the PCB. A team at Princeton University has discovered that they can train diffusion models to create them. They first trained a deep learning model to predict S-parameters from arbitrary metal structures. We then obtained the S-parameters and trained a diffusion model that transitions from noise to metal patterns with appropriate interference properties. The results are often more reminiscent of a 2D barcode than a pattern created by a manual designer.
Recent research findings may signal the beginning of a transition from analog automation to generative AI. More work needs to be done to assess whether using them is better than using human experience. This tool may prove even more useful when providing designers with a collection of options. Experience will guide your choices, understanding how sensitive you are to different situations. It also helps train new engineers by making it easier to consider the design space and various trade-offs.
The figure above shows one charge and image of the matching network layout produced by the Princeton diffusion mode of a single-port antenna.
“For many target requirements, there are often many ways to solve the same problem. For circuits such as band gaps and operational amplifiers, the plateau of solutions is often very large. For something like mmWave satellites, this may be an issue, but in other cases squeezing the last drop from the tea bag may not be a concern,” says Wong. He adds that in environments where correct tapeout the first time is important, experience with the existing topology is often more important than performance improvements that don’t have a significant impact on operations.
“But even if you’re ultra-conservative, you can see that these technologies are becoming very powerful, and those who adopt them effectively will have an advantage,” Wong concludes.
