South Korea and AMD Begin Domestic NPU Trials

South Korea’s Ministry of Science and ICT said it will work with AMD on a large-scale trial combining domestically made neural processing units (NPUs) with AMD central processing units (CPUs) and graphics processing units (GPUs). The ministry plans to build a reference system next year to lower barriers to entry in public and private markets. It announced the plan during a panel at Advancing AI World Tour and AI Dev Day Seoul 2026, held on the 4th at the Westin Seoul Parnas Hotel in Gangnam, Seoul. Representatives from the ministry, AMD Korea, Moreh and MangoBoost discussed strategies for national AI infrastructure self-reliance.

Lee Do-gyu, head of the ministry’s research and development policy office, forecast that inference would account for 60.0% of total AI workloads in 2030, with training at 40.0%. He said domestic NPUs with high performance per unit of power need to become mainstream to handle rising power use and operating costs for inference. The government is applying K-Perf, a set of performance indicators developed last year by a group of 15 technology demand and supply organizations, to validate hardware and establish commercial use cases. Lee said a company completed performance validation using K-Perf last month as a condition of purchasing NPUs worth 10 billion won.

A memorandum of understanding signed by the ministry and AMD in San Francisco in July specifies cooperation with South Korean NPU companies. Lee called the provision unusual and said it combines AMD’s open computing strategy with the government’s commitment to domestic production. Moreh CEO Kangwon Cho said software capabilities that are not tied to a particular chipmaker are needed, highlighting accelerator optimization that can run data centers on different semiconductors.

Cho cited Meta’s shift toward commercializing Llama and said open-source models from China could also become closed-source. Moreh said its subsidiary Motif is developing its own large language model (LLM) and building distributed training capabilities that connect thousands of accelerators into a single cluster. Cho said models can run reliably on AMD’s open-source software platform ROCm, and that software engineering determines performance differentiation.

MangoBoost CEO Jangwoo Kim said companies want to keep data and systems on site for security and sovereignty, but AI infrastructure is expensive to build. Costs include not only GPU memory but also high engineer salaries for optimizing open-source models, he said. He proposed a heterogeneous full-stack system that jointly designs accelerators, networking and storage to reduce computation. MangoBoost said it will build a testbed combining AMD CPUs and GPUs, domestic NPUs and South Korean system software in a single fabric, then use trial services to demonstrate better price-performance. AMD Korea CEO Junghee Jang said an open computing ecosystem involving government, industry and developers would establish technological choice and leadership in the South Korean market.

South Korea’s Ministry of Science and ICT said it will work with AMD on a large-scale trial combining domestically made neural processing units (NPUs) with AMD central processing units (CPUs) and graphics processing units (GPUs). The ministry plans to build a reference system next year to lower barriers to entry in public and private markets. It announced the plan during a panel at Advancing AI World Tour and AI Dev Day Seoul 2026, held on the 4th at the Westin Seoul Parnas Hotel in Gangnam, Seoul. Representatives from the ministry, AMD Korea, Moreh and MangoBoost discussed strategies for national AI infrastructure self-reliance.

Lee Do-gyu, head of the ministry’s research and development policy office, forecast that inference would account for 60.0% of total AI workloads in 2030, with training at 40.0%. He said domestic NPUs with high performance per unit of power need to become mainstream to handle rising power use and operating costs for inference. The government is applying K-Perf, a set of performance indicators developed last year by a group of 15 technology demand and supply organizations, to validate hardware and establish commercial use cases. Lee said a company completed performance validation using K-Perf last month as a condition of purchasing NPUs worth 10 billion won.

A memorandum of understanding signed by the ministry and AMD in San Francisco in July specifies cooperation with South Korean NPU companies. Lee called the provision unusual and said it combines AMD’s open computing strategy with the government’s commitment to domestic production. Moreh CEO Kangwon Cho said software capabilities that are not tied to a particular chipmaker are needed, highlighting accelerator optimization that can run data centers on different semiconductors.

Cho cited Meta’s shift toward commercializing Llama and said open-source models from China could also become closed-source. Moreh said its subsidiary Motif is developing its own large language model (LLM) and building distributed training capabilities that connect thousands of accelerators into a single cluster. Cho said models can run reliably on AMD’s open-source software platform ROCm, and that software engineering determines performance differentiation.

MangoBoost CEO Jangwoo Kim said companies want to keep data and systems on site for security and sovereignty, but AI infrastructure is expensive to build. Costs include not only GPU memory but also high engineer salaries for optimizing open-source models, he said. He proposed a heterogeneous full-stack system that jointly designs accelerators, networking and storage to reduce computation. MangoBoost said it will build a testbed combining AMD CPUs and GPUs, domestic NPUs and South Korean system software in a single fabric, then use trial services to demonstrate better price-performance. AMD Korea CEO Junghee Jang said an open computing ecosystem involving government, industry and developers would establish technological choice and leadership in the South Korean market.