사진 확대 Kang Ji-hoon, CEO of Puriosa AI (CRO), Kim Jang-woo, CEO of MangoBoost, head of PIM Artificial Intelligence Semiconductor Business at the Information and Communication Planning and Evaluation Institute, Lee Do-kyu, professor of KAIST, Cho Kang-won, CEO of the day after tomorrow, and Moon Kyung-sun, managing director of AMD Korea (from left) are taking commemorative photos at the “AI Semiconductor Dae-hwan” forum held at Maekyung Media Center in Jung-gu, Seoul on the 18th. Reporter Lee Choong-woo
The artificial intelligence (AI) infrastructure is changing. The era of filling data centers with one type of GPU is over, and the center of gravity is shifting to a different 機 structure that puts CPU, GPU, and NPU together. With the opening of the era of AI agents, it is not how well they make the model, but how cheaply and with little power they handle the flood of requests for reasoning.
Korea has become a country where only domestic companies can build an “open computing stack” from NPU to memory, data processing units (DPUs), servers, networks, and orchestration software through intensive support over the past seven to eight years. It is against this backdrop that the Ministry of Science and ICT and AMD signed a Memorandum of Understanding (MOU) to build an open AI infrastructure ecosystem in San Francisco on July 23. Deputy Prime Minister and Minister of Science and ICT Bae Kyounghun emphasized at a recent meeting of AI companies, “The AI market has not yet been created and we must pioneer it.”
Mail Business, along with the Ministry of Science and ICT, held an expert discussion on “AI Semiconductor Daejeon” at the Maekyung Media Center in Jung-gu, Seoul on the 18th and sought a solution to the global entry of domestic AI semiconductors. Lee Do-kyu, head of the Information and Communication Policy Office at the Ministry of Science and ICT, Moon Kyung-sun, managing director of AMD Korea, Kim Jang-woo, CEO of Mango Boost, Cho Kang-won, CEO of Moreno, Kang Ji-hoon, Chief Research Officer (CRO) of Puriosa AI, Oh Yoon-je, head of the PIM Artificial Intelligence Semiconductor Project at the Information and Communication Planning and Evaluation Institute (IITP), and Yoo Hoi-joon, professor of KAIST, attended.
사진 확대 Deputy Prime Minister and Minister of Science and ICT Bae Kyounghun (right) and AMD Chairman Lisa Su talk at the “AMD Advancing AI 2026” event held in San Francisco on the 27th of last month. Ministry of Science and ICT
-What does this cooperation between Korea and AMD mean.
Director Lee Do-gyu = The government first proposed a clause that links AMD’s CPU and GPU with domestic NPU and included it as the first cooperation task of the MOU. Until now, domestic AI semiconductors have been ‘subjects to support’ and have never been discussed as ‘components’ of global architecture. Securing it in writing is the biggest achievement.
Managing director Moon Kyung-sun = AMD’s three strengths include △ world-class memory competitiveness △ industrial-based △ start-ups to universities that will apply physical AI such as manufacturing, robots, mobility, shipbuilding, and disaster prevention. It is the first and very unusual case in Asia to decide to establish an excellent research center (CoE) as a pre-investment in the absence of large-scale prior hardware investment.
Kang Ji-hoon CRO= The way to grow the domestic AI semiconductor industry is not to create a closed ecosystem only with domestic technology, but to create an open system that customers can choose by connecting our technology with global CPU, GPU, and software ecosystems. Currently, AMD is discussing expanding the level of technical cooperation between the two companies to the data center system.
-How open do you think the window of opportunity is.
Professor Yoo Hoi-joon = It will be three years from now. AI semiconductors take at least two to three years from the start of design to mass production. This does not mean that the window will close in three years, but that the tickets for the ship that will arrive in three years should be cut off now. The data center’s accelerator replacement and expansion cycle is also three to four years, so if you miss the cycle now, the next opportunity will be in the 2030s. If you can’t make a reference in 3 years, there’s no next 30 years.
Director Oh Yoon-je = It is rare for the government to invest preemptively and continuously for seven to eight years. As a result, domestic NPU companies are making sales and exports. AMD is becoming virtually the only opponent that can challenge Nvidia’s stronghold, so now is the time when each other’s needs are right.
-So what’s holding back.
CEO Kim Jang-woo = Nvidia and AMD are also trying to set up the entire system these days. When the system is tied up as a whole, vendor dependency is created, and no matter how good the parts are, there is no place to put them. You should do it when AMD opens the door. True Sovereign AI is possible only when a Sovereign AI ‘system’ is secured.
CEO Cho Kang-won = However, since there are A, B, and C in Korea, the approach to tying them all together and making a big one at once does not work well. Even small integrations must be tried and failed in several combinations to define the interface and the maturity of element technology. With such accumulated experience, greater integration is possible in the future.
-How will next year’s demonstration be successful.
Kim Jang-woo’s representative = Sung plaque depends not on “what chip to put in,” but on “who draws the architecture.” It is meaningful only when it is verified what performance and cost the NPU performs in the actual data center architecture rather than how much performance it performs alone.
Professor Yoo Hoi-joon = The demonstration should be an actual business, not a demonstration. It should be a real service, not a benchmark, and it becomes a reference only when you receive real traffic and experience failure and recovery. It should be deployed as the main calculator to disclose performance per power and total cost of ownership (TCO) in numbers that can be verified by third parties.
Kang Ji-hoon’s CRO= demonstration should not end as a pilot project, but should lead to actual commercial services. Semiconductors, servers, networks, clouds, and demand institutions must participate with common business goals to demonstrate the entire ecosystem.
General Manager Oh Yoon-je = We need to go step by step. The first stage is to verify the software linkage between AMD’s CPU and GPU and domestic NPU, the second stage is to be a pilot that applies to actual service workloads, and the third stage is to disclose the results as objective indicators and reflect software assets to open ecosystems such as ROCm andUALink.
Director Lee Do-gyu = The government is preparing for large-scale demonstration as a new budget project next year. The spread does not occur because the government buys it, but only when the private sector wants to buy it. What the government should do is to buy time and provide a basis for judgment until then.
-Where should domestic AI semiconductors be in 3 years.
Professor Yoo Hoi-joon = In 2029, ‘If power and cost are a problem, review Korea’s NPU first’ should be a natural option at any data center in the world. At the same time, we need to prepare chips after chips. The next generation is a memory-driven computation structure, followed by PIM (intelligent memory), followed by neuromorphic. PIM is virtually the only track that is structurally advantageous for Korea, which has both the first and second largest memory companies.
Kang Ji-hoon CRO= The goal is to become a solution that is naturally reviewed when building a large-scale reasoning infrastructure anywhere in the world. Compared to Nvidia GPUs under the same conditions, its NPU Renegade (RNGD) has handled 1.8 to 2 times more concurrent users per power, satisfying service-level targets per user. However, the developer problem is not solved by ‘use it because it is made in Korea’.
CEO Cho Kang-won = AI Data Center will go with a heterogeneous architecture that mixes various chips. Because one chip can’t always be the best for every task. When global companies build new data centers, they should aim for the direction of being naturally included in the list of chips.
Managing Director Moon Kyung-sun = AMD expects Korean NPU companies to establish themselves as key players in global AI infrastructure, not as “domestic substitutes.” If the government, industry, and academia cooperate, Korea can not become an AI consumer, but a country that exports AI platforms and AI semiconductors.
Director Lee Do-gyu = Taking this opportunity, Korea’s AI semiconductors should take a leap forward as a key component of global architecture. The MOU is a declaration to the end. Making a line written in the document into a system that actually works, that’s the remaining task.
Intelligent Memory (PIM)
It is a next-generation AI semiconductor that integrates computational processor functions inside memory semiconductors that only stored data, increasing the speed of computational processing and drastically reducing power consumption.
neuromorphic
It is a next-generation semiconductor technology that mimics the structure and operation of human brain neural networks, and it is very advantageous for AI inference and sensory data processing by performing computation and storage at the same time.
[Reporter Kim Daegi]