Software developed to crowdsource electronics design to improve machine learning

23-11-2018 | By Rob Coppinger

In a year’s time electronic integrated circuit designers will be able to download a programme to help them design and which will send back data to its developers to improve machine learning for microchip layouts.

The crowdsourcing of designers and designs for machines to learn from is an attempt to reduce the costs of microchip development. Creating the chip layout design and then turning that into manufacturing data for the lithographic process that makes microchips takes a long time. The cost of this design for manufacturing process is not decreasing and as microchips get more transistors and more complex the cost is going up. The goal is to create a community of open source users that use the free design software, share information among themselves and send data back to the application’s developer for improving the machine learning.

“When we use machine learning we learn from existing designs,” says Sachin Sapatnekar, a University of Minnesota professor of electrical and computer engineering. “Out there are a large number of designs behind IP [intellectual property] barriers, but if someone could run our software and run it on their design, they could feedback some information that could improve the overall [machine learning] programme.”

Circuit design

A transceiver integrated circuit die shot used in transmitters and hand held computers. By Kawe Mazidjatari - Own work, CC BY-SA 4.0, Link

Sapatnekar will lead the work at the University which will have Texas A&M University and semiconductor giant Intel as partners. He expects to have the first version of the free software ready for downloading by December 2019 from file sharing sites such as Github. The first version will work for a subset of circuits, later versions will be more capable.

Another reason for the complexity of modern-day chips is that they are typically hybrids of digital and analogue designs. This is because input devices, such as a microphone, produce an analogue signal. The researchers hope an open-source intellectual property ecosystem with an automated physical layout generator, the machine learning, will see the physical design of such hybrid electronics completed within 24 hours; not weeks or months. Another goal for the project is to create software that makes it easier for new entrants into the microchip design market which is dominated by a few companies such as Intel and Nvidia.

Sapatnekar expects the early microchip designs produced through this ecosystem of users and machine learning to be less powerful than existing chips. “If you want to do this in a day [create the manufacturing design files],” he explains. “You have to give up a bit on performance, but get the functionality going and improve the performance later.” His view is that performance will improve over time.

The University of Minnesota work is part of a four-year, $5.3 million Intelligent Design of Electronic Assets (IDEA) programme grant from the United States’ Department of Defense’s Defense Advanced Research Projects Agency. The University of Minnesota is one of 11 lead universities or companies to receive funding from the IDEA program.

By Rob Coppinger

Rob Coppinger is a freelance science and engineering journalist. Originally a car industry production engineer, he jumped into journalism and has written about all sorts of technologies from fusion power to quantum computing and military drones. He lives in France.