Machine Learning Is a Key Component of Future System Design and Platform Creation

Elias Fallon is the engineering director of cadence design systems, an industry-leading provider of electronic design automation technology. He led his custom IC R & D team and electronic design automation (EDA) product team in project development.

Elias Fallon, who is responsible for the EDA research project of applying machine learning and deep learning technology to computing software, has a unique insight into the future development of the electronic design industry.

He pointed out that the next leap in the design of semiconductor chips and their peripheral systems will come from the integration of EDA computing software tools and processes, as well as the application of larger-scale machine learning / deep learning (DL) technology and multi-core computing. The current wave of innovation in artificial intelligence and machine learning begins with the improvement of GPU computing power and the development of methods by design engineers to speed up deep neural network training. Machine learning / deep learning will play a key role in the design of the next generation platform, so that emerging technologies (including 5g, super large-scale computing and other technologies) can be widely used.

For Fallon, the fun is to solve some non deterministic polynomial (NP) problems and integrity problems at each stage of the design and verification process. Fallon and his team developed software that design engineers use to design, simulate and verify integrated circuits, packages, circuit boards and systems. The design challenges they face are so difficult and complex that they don't find the best solution in a certain time. By definition, the verification challenge is a problem that has never been encountered before. Fallon and his team have developed a variety of complex algorithms and software to provide the best solutions. These innovations have promoted the improvement of design productivity for customers.

Numerical solver, Boolean satisfiability solver, adaptive meshing, computational geometry and iterative improved optimization algorithm are all examples of computing software. Computing software algorithms require EDA software engineers to determine how to best apply the algorithms to current design challenges and how to present various meta parameters, controls and commands to users in design terms. The trend of next-generation design is to increase complexity in system design and verification, which will require the addition of computing software "tools" in EDA toolbox to achieve a leap in design productivity.

Over the past six months, people working remotely from home have benefited from the progress of cloud computing, chip optimization and the Internet. In a virtuous feedback cycle in the electronics industry, computing software provides help for electronic design, and the design team will also benefit from future innovation. 5g, ultra large scale computing and other technology drivers require a lot of innovation in chip, packaging, circuit board and system design to create various possibilities for the development of electronic technology in the future. The machine learning function through example learning is a new computing software tool, which lays the foundation for the next round of innovation of designers.

Computing software has achieved great growth in productivity improvement and solving the complexity challenges of electronic system design. When cadence's solution solves the challenges of the previous generation, the design of the next generation is more complex. The increasing complexity of the system also brings the complexity of the design and verification process. This complexity has become an obstacle to changing processes to adopt new best practices or automation. Each tool or step in a complex process (adding new options, commands and functions) requires users to understand, evaluate and check its adaptability in the overall process before it can be adopted. Have the ability to deploy machine learning, learn design practices from users through examples, and allow EDA software engineers to develop systems to transform machine learning design practices into options in new tool processes, thus accelerating the adoption of innovative design processes.

For example, analog circuit designers will understand the devices that need to be matched in circuit design and layout based on previous experience, but to adopt automation technology in the design process, additional constraints and specifications need to be added. Machine learning model can learn the best practices of those designers from the completed design, and accelerate the whole design process in a customized way for each designer or design team. Innovative systems design companies do not share their designs or machine learning models trained from them with other companies. Therefore, the training of learning the design practice of machine learning must be carried out at the user. Machine learning will become a key tool in EDA computing software toolbox, which looks different from many SaaS based machine learning products in other industries.

In addition to learning design practices through examples, the most common use case for machine learning in EDA is to predict future process steps. When all these non deterministic polynomial (NP) problems are put together, it is difficult to fully predict the impact of the current process step results. The most common example is understanding routability when optimizing layout. The layout usually determines the location of each component and aims to minimize area / cost and wire length. Routing establishes a connection for each signal between each component. These components may be components on a PCB or modules or transistors on a chip. Over the years, EDA engineers have developed many heuristic methods to improve wire length and routing capability while optimizing circuit layout. However, because routing and placement is a non deterministic polynomial (NP) problem, it is computationally infeasible to try each option, and the existing heuristic methods may miss many nuances of routability.

By using machine learning model, taking layout as input and routing score as output, we can potentially create a richer and faster solution. EDA tool flow can generate multiple candidate layouts, route each candidate layout, and use the routing score as a marker to train the machine learning model. Similarly, any flow that generates many design candidates and results by running a complex EDA flow can build a model to predict the results of future steps from the input of previous steps. This provides a greater ability to find better solutions in complex design space.

Machine Learning Is a Key Component of Future System Design and Platform Creation 1

future system design and platform creation related articles
Back to School Help!Read Details! Get 10Points For1 with Best Outfit!?
Design Scheme of Multi-channel Voltage Measurement Based on Stm32
How Does Temperature Affect the Speed of Sound?
What Is WiFi? Just Tell Me What It Is, and What It's Used For.?
Which Is Better, WiFi Or ZigBee Wireless Technology

KingBird Home Furniture