Devgan plans for optimization through machine learning

By Chris Edwards |  No Comments  |  Posted: July 12, 2022
Topics/Categories: Blog - EDA  |  Tags: , , , , , ,  | Organizations:

Anirudh Devgan wants people to stop calling everything AI. But for the president and CEO of Cadence Design System, machine learning across the tool portfolio seems inevitable.

In his keynote at DAC in San Francisco today (July 12th), Devgan said, “It’s not good for the field to call everything AI. For me, much of it is more a pattern-based algorithm: a data-driven approach.”

The company has launched a couple of tools in its Cerebras lineup that are based on reinforcement learning. Devgan said he sees this the company leaning more on approaches based on reinforcement learning, particularly those based on gradient-free training strategies in contrast to the stochastic gradient descent strategies used by the convolutional deep-learning models.

In these tools, there is still an analytical algorithm at the heart, which is combined a machine-learning model. “That can provide true value,” he claimed. “EDA historically has focused on a single run. There is no mathematical way to transfer the knowledge gained from one run to another. This is where the data-driven approach is useful.”

With machine learning used to build models from a series of runs, a tool can generate more effective settings that are most likely to yield the best results. This is the approach that largely lies behind the Cerebras products.

“Even our own team in the beginning was not convinced it can beat a human approach. But the results are spectacular. We can get 10x and more productivity benefits with better results,” Devgan claimed. “Instead of the designer spending time on mundane tasks such as choosing tool options, they can make decisions at a higher level.

“With gradient-free learning, f(x) can be any function. You don’t need to find [the derivative] f’(x),” Devgan noted, which provides the option to apply the multi-run optimization tactic to many different domains.

“We are also applying this to PCBs with very encouraging results. An equal amount of optimization has to be applied to package and PCB design as for IC because they will all be critical going forward. And doesn’t matter what f(x) is to the AI-driven system-optimization flow. You can put CFD in the core,” Devgan added, pointing to the technology for thermal analysis that the company is pulling in through its acquisition of Future Facilities.

Less clear is the role of machine learning in verification. “It’s a never-ending problem. You never know when you are truly done. AI can play a role here. This is where we will need a lot of help from the university folks.”

Comments are closed.


Synopsys Cadence Design Systems Siemens EDA
View All Sponsors