Machine Learning Electronic Design Automation: Unlocking New Designs
The scope of machine learning in EDA.
How machine learning works with designers to improve designs.
The intricate relationship between machine learning, high-level synthesis, and design space exploration.
Machine learning implementations exist on a spectrum, with more comprehensive solutions requiring more intensive and novel models
Electronic design automation software has greatly reduced the barrier of entry for circuit design and increased the speed of development by providing users with powerful toolsets. In much the same way, machine learning is positioned to act as the new revolution in EDA software solutions by offering the ability to automate large sections of the design and layout process. While there are still some significant shortcomings to grapple with, the field is rapidly growing with new and innovative techniques that solve design challenges more efficiently. Like the math and programming that are driving the relatively nascent development, machine learning electronic design automation offers the promise of abstracting design, leaving the design work that is largely computational to the computer and freeing the designer to best meet the challenges of high-level design.
The Levels of Machine Learning Electronic Design Automation Functionality
The breadth of the solutions across a wide amount of industries is one of the key advantages of machine learning from electronic design automation software. The ability of implementation is also staggering–countless methods are currently in use to help lead to quicker turn times on development. For machine learning in electronic design automation, there are four aspects of uses that are leading to reduced turn times and improved results:
Decision making - Here, a model is trained to better search through available toolsets and algorithms to replace brute-force searching or unassisted searching. The concept is that leading the model with machine learning helps make its active decision-making equally comprehensive while being quicker to hone in on the methods that will lead to increased efficiency.
Predictive ability - Models can analytically view completed designs to help guide decisions on new designs. By leveraging this database, layout designers can see a huge gain in efficiency by having a machine learning toolset gauge design pathways without having to invest exceedingly large amounts of time into the actual design process.
Optimization - The design space exploration, or DSE, is navigated solely by a machine learning model and whatever data science theories are providing guidance. Generally, this takes the form of regression-based models that are able to target an error function (absolute, root mean square, or an entirely different function) to actively form the layout to the tangible functional results. This occurs in two steps. First, input and output values are taken into consideration that at the most concrete level provide a gauge of the overall functionality of the circuit. Tracked values could clock periods (to determine time), memory allocation for bandwidth or storage, and more. With the established values, the circuit can receive some internal level of feedback, although this is less a “conscious” effort and more of an ongoing calibration to hit a moving target.
Full automation - The apex of machine learning electronic design automation. In this setup, data can be synthesized in real-time using machine learning heuristics, such as deep learning or reinforced learning. How this establishes itself in practice is a system that is able to take into account present data points, alongside some sort of critique/incentive framework operated by two complimentary sub-routines, to make gains in optimization at an appreciable rate to human operators. The ideal deployment of an automated system still takes into account an engineer or layout designer’s expertise in the field to oversee, review, and finalize the design once the automated layout process is completed.
The Symbiosis of Operators and Designers
As evident above, machine learning in electronic design automation has several stages of implementation that reflect a gradual increase in the share of the automation compared to the overall design process.
Full automation, which may seem the stuff of science-fiction, is gaining adoption today and has even seen usage in cutting-edge commercial applications. Properly structured, the computational ability is able to solve problems on time scales comparable to that of expert designers, and in some cases much more expeditiously. Automation can’t completely replace the design expertise of a layout designer–it is limited only to circuit permutations based upon the dataset used to train it. In other words, machine learning is only equipped to handle the designs it has experience with.
More sophisticated machine learning models, however, can blur this line by generalizing the design constraints of a board. By focusing on concepts such as power draw, area, bandwidth, and other similar, well-defined values, machine learning software that is properly trained is able to predictively problem-solve specific datasets it has no experience with. This occurs without any level of hard-coding either; the system is able to deduce from prior incentives the best route forward for the layout, depending on what aspect or aspects need to be optimized.
How High-Level Synthesis and Design Space Exploration Mold EDA
High-level synthesis, or HLS, is able to automatically convert between high-level programming languages to hardware description languages like VHDL to implement the operation of logic gates at the schematic level into an FPGA for debugging purposes. HLS has existed prior to machine learning, but gains have been made to increase the predictiveness of the translations as well as improvements to current DSE algorithms, including rethinking the approach to historical DSE in its entirety. More specifically, this predictiveness manifests as a result estimation, essentially a target the system has to hit that balances accuracy against efficiency. Result estimation will occur in two steps:
Preparation - The machine learning method must be not only trained on a wide dataset to prepare for the different circuits it will encounter but also invoke different timings and FPGA targets. The data must proceed sequentially through the entire C-to-bitstream data flow.
Modeling - Regression-based modeling and analysis are performed on the dataset to determine resource usage and timing. Different machine learning models are applied in a synergistic manner to the highest accuracy of the estimation.
For DSE, a variety of models are employed to improve longtime standard methods. Predictive models are broaching this problem in separate ways: some approach the accuracy as the goal of the model, while others address the inaccuracy of the learning tool directly. Models like transductive experimental design (TED) look to balance not only representative members of the dataset but also the outlier cases that may be harder to predict.
For any level of design needs and machine learning electronic design automation integration, Cadence’s PCB design and analysis software tools are well-suited to meet the needs of today’s designs as well as unforeseen future developments.
Leading electronics providers rely on Cadence products to optimize power, space, and energy needs for a wide variety of market applications. If you’re looking to learn more about our innovative solutions, talk to our team of experts or subscribe to our YouTube channel.