South Korean Government to Start Localizing Major Infrastructure for Artificial Intelligence

Feb 25, 2016

South Korean Government is going to localize major infrastructure that can implement artificial intelligence (AI). Its goal is to reduce technical difference with other advanced countries in AI area that is one of major key words of ‘4th Industry Revolution’.

According to South Korean Government on the 24th, Ministry of Science, ICP, and Future Planning (MSIP) is going to invest $5.65 million (7 billion KRW) by 2019 and develop high-performance computing (HPC) system for Deep Learning. It is going to localize major infrastructure that can implement Deep Learning, which is one of AI areas.

Deep Learning is AI technology that enables computers to learn on their own and has them recognize and analyze data. Google’s AI ‘AlphaGo, which has challenged Lee Se-dol who is the world’s Go champion, is also based on Deep Learning algorithm. ‘Robot Sound’, which is a movie that had received attention for showing communication taking place between a human and a robot, is also based on Deep Learning.

If Deep Learning is to be implemented, computers have to go through ‘Model Training’ that enables them to learn on their own. This training helps computers to learn languages, images and others repeatedly. Model algorithm and hardware (HW) with large capacity and high performance are must for computers to absorb tremendous amount of data quickly. Google, Facebook, Baidu and others have secured such technology with their own parallel distributed processing framework and large data centers.

Concept diagram of Distributed Parallel Processing for large-sized Deep Learning model (Reference: IITP) <Concept diagram of Distributed Parallel Processing for large-sized Deep Learning model (Reference: IITP)>

South Korean Government is going to develop Deep Learning’s major software (SW) called ‘Multi-Node Distributed Processing SW Framework’ for the first time in South Korea. Model Training that was done by a single server is going to be processed by multiple servers. South Korean Government is targeting 2 hours and 9 hours for large amount of image recognition time and voice recognition time respectively and 60 milli-seconds for performance of Deep Learning Training Accelerator. Image and voice recognition time for Baidu and MS, which currently have world’s best level of time, are 8.6 hours and 39.27 hours respectively. South Korean Government’s ultimate goal is to reduce them by sixth.

South Korean Government is also going to develop a server that supports training. It is highly likely that this server is going to be FPGA-based (Field Programmable Gate Array) Deep Learning Server and not widely-used x86 server. FPGA is a processor that was designed to only perform particular programs and is used for places that require fast processing speed for Big Data, search engine and others.

“There weren’t any instances when Multi-Node Distributed Processing technology was commercialized in Deep Learning field.” said Professor Lee Jae-jin of Seoul National University’s Computer Engineering Department. “It will be valuable to secure FPGA technology that is being expanded towards Big Data and AI areas.”

AI, which is one of major key words of 4th Industry Revolution, is being recognized as an innovative technology. Many governments and businesses in the U.S., Japan and others are working towards securing this technology. Currently South Korea is behind 2 to 3 years compared to other advanced countries.

It is meaningful that development of HPC technology for Deep Learning that South Korean government has prepared bases for securing national competitive edge.

More than anything else, this business is good news for small businesses. When development of this technology is finished, South Korean Government is going to provide this technology in IaaS (Infrastructure as a Service) method. Small businesses that are located in Pankyo Creative Economy Valley will receive this technology first. They won’t have to purchase expensive GPU servers for Big Data, Deep Learning and others or use HPC Cloud services from AWS (Amazon Web Service), Google and others. They can just now use HPC Cloud that is provided by South Korean Government at an inexpensive price.

“Small businesses that did not have system that can analyze data even when they had data now can use HPC resources with an inexpensive Cloud method.” said Park Yoo-mi who is a senior researcher for Electronics and Telecommunications Research Institute (ETRI). “This business will also be a big help for startup companies with competitive edge that are looking to develop business models that use Deep Learning.”

Staff Reporter Jung, Yongchul | jungyc@etnews.com

▲In Deep Learning Training acceleration performance, South Korean Government is trying to reach 60 milli-seconds and currently America’s NVIDIA has the world’s best time at 313 milli-seconds.
▲In large image recognition performance, South Korean Government is trying to reach 2 hours and currently China’ Baidu has the world’s best time at 8.6 hours.
▲In large voice recognition performance, South Korean Government is trying to reach 9 hours and currently America’s MS has the world’s best time at 39.27 hours.
▲In expandability of Deep Learning Parallel Model, South Korean Government is trying to reach 12 servers and currently China’s Baidu has the world’s best number of servers at 8.
<▲In Deep Learning Training acceleration performance, South Korean Government is trying to reach 60 milli-seconds and currently America’s NVIDIA has the world’s best time at 313 milli-seconds. ▲In large image recognition performance, South Korean Government is trying to reach 2 hours and currently China’ Baidu has the world’s best time at 8.6 hours. ▲In large voice recognition performance, South Korean Government is trying to reach 9 hours and currently America’s MS has the world’s best time at 39.27 hours. ▲In expandability of Deep Learning Parallel Model, South Korean Government is trying to reach 12 servers and currently China’s Baidu has the world’s best number of servers at 8. >

Interpretation & Translation_Service Center

Refund Help Center