[Herald Interview] ‘Semiconductors still have so much room to grow’

DAEJEON — In 2000, professor Yoo Hoi-jun of the Korea Advanced Institute of Technology predicted that cellphone screens would grow larger and users would play games on the phones.

To prove his prophesy, he confidently showcased demos of mobile chips to tech titans like Samsung, Nvidia and Sony at the time, though to lukewarm responses.

In 2006, Yoo came up with even more advanced argument: artificial intelligence would be the future.

The expert on electrical and electronics engineering was inspired by the book “Astonishing Hypothesis” by Nobel laureate Francis Crick, who co-discovered the molecular structure of DNA.

Yoo was reading the book at a cafe in San Diego after presenting a paper on a neural network system that fastens and improves processing graphic data on mobile devices at an academic conference in the Silicon Valley.

“Inspired by the book, I began developing brain-like neural network chips, believing that the structure of mobile chips would resemble human brain as AI becomes a new norm,” said Yoo in an interview with The Korea Herald.
Since then, the professor’s team at KAIST has concentrated their efforts on AI chipsets for mobile devices, and it has won global attention as the team introduced an integrated deep neural network system — the deep neural processor unit, or DNPU — at the Hot Chips 2017 conference in August.

In the chip structure, the convolutional neural network is used to support vision recognition and processing, while the recurrent neural network is able to recognize time varying entities and support generative models.

By combining both CNN and RNN, a system can recognize time-varying visual entities, such as action and gesture, and support image captioning, according to Yoo’s paper.

“We focus on reducing power consumption and processing graphic data at real time, when a chip runs in mobile devices, drones and automobiles,” the professor said. “Since the beginning, we have bet on mobile embedded chips.”

Google’s AlphaGo that ignited the current AI boom across the globe has adopted Nvidia’s graphic processor unit, which has turned out to be a huge power guzzler.

“In the coming February, we plan to showcase demos of more power-efficient and simpler chips that can recognize human gestures,” he said.

With those embedded chips, the professor forecasts a future with “friendlier” smartphones.

“I don’t think current devices are intelligent enough,” Yoo said. “Until now, devices themselves haven’t changed much, while data has been piling up and memory has been running out.

“I expect smartphones to become more customized to users with on-device deep neural networks that support operations of applications just like a human brain does.”

Contrary to a government-level initiative to raise software power, Yoo stresses that Korea still needs to invest more in hardware-like chipsets in order to retain the world’s No. 1 position.

“Samsung Vice Chairman Kwon Oh-hyun’s message that Samsung needs some breakthroughs should be read seriously,” the professor said. “Korea became a chipset powerhouse thanks to government policies, conglomerates’ swift decision-making and elite workforces (about three decades ago).

“There are (now) no government-led research projects on semiconductors, and we are lacking resources to raise young experts,” he said. “The government is sending a message to young students that software is the future and that we are done with the chips.”

According to the professor, the number of students studying semiconductors has reduced by a third compared to in the early 1980s, while outstanding talents now tend to shun the field.

“Samsung’s Kwon also has been mentioning the need to address the workforce shortage, which can’t be resolved at a company level,” he added.

The future of the semiconductors industry can take shape through AI, biomedicine and quantum computing, Yoo said.

“Finding new materials and principles for making chips can provide room for innovation in the industry,” he said. “Making quantum devices and organic devices for computing is also being suggested by experts around the world.”

“Development of high-performance system on chips for AI and biomedicine is also part of the future,” he added.

Processor in memory, or PIM, can be a new lucrative market in the AI era, Yoo suggested.

“PIM is a product type that integrates memory chips with a central processor unit, which is mostly sold by Intel,” he said. “Samsung once had attempted to sell similar memory-embedded processor in the late ’90s, but the market was too small at the time.

“But PIM can be popular in the near future, as demand for processors with high memory capacity continues to rise to run AI applications.”