Collaborative research between ten university teams and government-supported research institute
Concentrating on development of post-Si semiconductors
Demonstrated ‘neuromorphic chip’ that mimics how the brain functions
Director of the research institute announces, “Semiconductor chips emulating neural networks and operating on low power will be developed within three years.”
SNU ECE professor Kiyoung Choi is explaining the neuromorphic semiconductor at the KIST Post-silicon Semiconductor Institute on the 23rd.
In the KIST (Korea Institute of Science and Technology, President Lee Byung Gwon) L3 building located in Hwarang-ro, Seoungbuk-gu, Seoul, on the 23rd. Amidst rising concerns that even our key industries are facing predicament with China’s ‘semiconductor uprise’, ‘KIST Post-silicon Semiconductor Institute’ is pushing forth technological development to out-distance China. Researchers are concentrating on the development of ‘post-Si semiconductors’, ‘AI neuromorphic semiconductors’, and ‘quantum computing’. Specifically, AI semiconductors that imitate neural networks and are capable of self-learning are being investigated by ten teams including KAIST(Korea Advanced Institute of Science and Technology), POSTECH(Pohang University of Science and Technology), UNIST(Ulsan National Institute of Science and Technology), Kookmin University and UCI(University of California, Irvine).
On this day, the research center demonstrated an AI semiconductor prototype that mimics neural network(Neuromorphic Chip. Neo²C) and is capable of self-learning without manual inputs. This chip efficiently utilizes memory semiconductors, specifically the SRAM block. Therefore, its commercialization after 2021 is expected to bring a synergy effect with Samsung Electronics and SK Hynix. It is anticipated that domestic companies will expand their market to non-memory semiconductors, which is currently in need of improvement. Earlier, Intel was the first to release a neural network mimicking semiconductor prototype in September, and China is investing massive amounts into AI and semiconductors, and thus, rapidly catching up. We are also engaging in the competition through collaboration between national research and development institutes and universities.
Neuromorphic semiconductors use AI technology based on hardware which is modeled on how the human brain processes and stores data. A total of 1.2 billion won was invested over six years beginning from 2016 into the three research areas required to imitate the brain’s neural network: the development of algorithm and architecture, circuits, and system of circuit and imaging neural networks.
In the neuromorphic chip, ‘neurons’(nervous cells) and ‘synapses’(connecting gaps between neurons) similar to those of the human brain were implemented digitally. It attempted to copy how the brain, through interactions between neurons, controls the intensity of synapse connections as it processes data by transmitting electric signals back and forth. KIST Ph.D. Jaewook Kim explained, “Unlike the DNN(Deep Neural Network)-based hardware accelerator chip that is already widely used, this is based on SNN(Spikin Neural Network), which imitates the human brain. It has 1,024 neurons and 199,680 synapses within an area of 16㎟, which makes online unsupervised learning possible in real time.” On this day, the research institute demonstrated how the neuromorphic chip recognized a bar, its angle, and selectivity according to variations in synaptic weights.
Joonyeon Chang, Director General of KIST Post-Si Semiconductor Institute, said, “Our goal is to develop imitated network semiconductor chips that can operate on low power and are capable of self-learning by 2021. If this is commercialized, it will motivate new growth not only in memory semiconductors but in ICT(Information and Communication Technology) fields in general.” Our technology is still evaluated to be half the level of Intel, USA, but because no entity has succeeded in commercialization, we have a chance.
With the Neuromorphic Chip, changing functionalities is possible. Its neural network structure and algorithm can be programmed in any manner. Its operation has been optimized by efficiently utilizing the area of the SRAM memory block used in the chip. As it can check the input and output based on actual neural networks without additional calculations, it can operate more efficiently and on lower power when compared to software AI semiconductors.
Current AI technology requires an immense amount of power for data processing and storage based on deep learning network. In 2016, during the Go match between Google’s AlphaGo and Sedol Lee (9 dan rank), AlphaGo consumed 170kW of power while Mr. Lee used only about 20W. KIST Ph.D. Jongkil Park compared the two, saying, “Whereas the brain is perfectly efficient in using energy for calculation and memory, software-based AI such as AlphaGo consume excessive power to run complex applications because it has to drive high performance semiconductor devices(CPU and GPU(Graphic Processing Unit)).”
This is why the implementation of neural network mimicking semiconductors is actively researched around the world. In the U.S., IBM announced its development of a chip called TrueNorth in 2014, but the chip could not train on its own. The prototype Loihi developed by Intel last year is considered to be the first semiconductor chip to learn and recognize visual information. It contains 1,072 neurons and 130 million synapses within 60㎟ and is more complex than a lobster’s brain. The Loihi has 128 cores while our neuromorphic chip has only one, so we have a lot to make up for. “Intel has provided the Loihi to universities and research centers, leading to further progress and development. We also hope to contribute to research innovation by supplying the neuromorphic chip to companies, universities, and research centers,” said KIST Ph.D. Joon Young Kwak.
By inviting participation and thus accelerating not only the development of semiconductors but also the merging of AI and neuroscience, the research institute plans to develop an original AI neuromorphic chip chat can learn on its own. Director General Joonyeon Chang broke into a wide smile, commenting, “In the future, we will expand the core and collaborate with companies to release a neural network inspired semiconductor that has one million neurons and two hundred million synapses by no later than 2021. We also plan to develop algorithms that can be applied to everyday life, such as in home robots or autonomous vehicles.”
Translated by: Jee Hyun Lee, English Editor of Department of Electrical and Computer Engineering, firstname.lastname@example.org