KAIST Develops New AI Chip Capable of Deep Learning

The researchers, led by professor Yu Hoe-jun at the Korea Advanced Institute of Science & Technology, announced on Monday that they had revealed findings from their study last week at Hot Chips, a symposium on high performance chips which was held in San Jose. (Image: IITP)

SEOUL, Aug. 28 (Korea Bizwire) – A South Korean research team has developed a new artificial-intelligence chip capable of deep learning that could be used in future mobile phones, robots and drones.

The researchers, led by professor Yu Hoe-jun at the Korea Advanced Institute of Science & Technology, announced on Monday that they had revealed findings from their study last week at Hot Chips, a symposium on high performance chips which was held in San Jose.

The semiconductor chip, also known as a deep neural network processing unit (DNPU), is capable of a number of AI-based tasks including object and action recognition as well as image captioning, which explains images in text, all of which fall under the category of deep learning.

As part of machine learning, deep learning is technology that allows computers to learn things on their own, including recognizing the owner of a device, and turning images into text.

The research team claims the energy efficiency of their DNPU is up to four times greater than the tensor processing unit which was used in Google’s famous AI computer program AlphaGo, which plays the board game Go.

A South Korean research team has developed a new artificial-intelligence chip capable of deep learning that could be used in future mobile phones, robots and drones. (Image: IITP)

In addition, unlike a TPU, which relies on data center servers to process deep learning, the DNPU chip can complete the task from devices like mobiles, robots and drones.

Hot Chips, where the findings from the study were first unveiled, is an academic platform where IT giants like Intel, Google and Microsoft have revealed groundbreaking advances.

This year, KAIST and Harvard University took part in the symposium. 

Ashley Song (ashley@koreabizwire.com)