Xilinx infuses EDA processes with machine learning
Source – https://www.fierceelectronics.com/
Xilinx today introduced the latest evolution of its Vivado Design Suite software, Vivado ML Editions, to help developers in the electronic design automation (EDA) sector leverage machine learning algorithms and techniques to bring greater efficiency to the design process.
The EDA industry is dealing with trends toward lower-nanometer architecture with increasinging complexity and numbers of transistors. It has “decades of data and engineer know-how” on design processes, but processes can be slowed by the massive numbers of iterations required to get the designs right and the lack of Quality of Results that come through those iterations, according to Nick Ni, director of marketing, Software and AI Solutions at Xilinx.
Companies in EDA also don’t have much of their own expertise in machine learning (ML), a technology that could help these processes become faster and more efficient. That’s where Xilinx can step in, having built up a lot of its own ML expertise in recent years, in part through the 2018 acquisition of China’s DeePhi Tech, a company focused on ML and optimization of neural networks, Ni said.
Jim McGregor, principal analyst at Tirias Research, agreed, telling Fierce Electronics via email, “Xilinx is in a unique position in that it has a large history of information and works with most of its customers on designs. Combined with Xilinx’s expertise in ML, the company has a unique ability to implement ML for the benefit of customers using its technology.”
Vivado ML Editions, available now, leverages machine learning-based algorithms to accelerate design closure, Ni said, and software features such as logic optimization, delay estimation and intelligent design runs, which drastically reduces timing closure iterations. ML Editions also includes an Abstract Shell feature which decouples compilations from platform interfaces. Ni said that’s important because it helps speed up compilation, while also allowing EDA companies to keep their intellectual property under wraps as they expose some design aspects to customers and partners during the design process.
These capabilities combine to allow ML Editions to deliver 5x to 17x faster compile time, and average breakthrough quality of results (QoR) improvements of 10% on average and all the way up to 50% on complex designs, compared to the company’s existing Vivado software.
“Without something like intelligent design runs, you are doing a lot of manual runs,” Ni said. “With IDR it’s a push-button process,” which allows those iterations to be completed, for example, in a day instead of a week or more with manual processes.
National Instruments is one of the first users of ML Editions, and in a statement, Robert Atkinson, the company’s principal hardware engineer, described the IDR functionality as “a game changer. By offering a push-button method for aggressively improving timing results, it generates QoR suggestions that bring maximum impact and deliver expert quality results with a reduction in user analysis – especially for tough-to-close designs.”
Kevin Krewell, also principal analyst at Tirias, called ML Editions a “really good example of applied machine learning in design tools. Using machine learning, the tools can converge on an optimal solution many times faster than traditional heuristics and with a better result.”
McGregor added that while the EDA sector is just beginning to use ML technology, usage will continue to grow. “You will eventually see ML throughout the semiconductor design, manufacturing, and even lifecycle management process,” he said. “Even though it is piecemeal now, adding ML to both hardware and software tools will provide a significant benefit and provide a competitive advantage for both the technology provider and the customer.”