AMD's Korea test lab is a bet that inference stops being a GPU-only purchase

AMD will establish an AI Center of Excellence in South Korea by the end of this year, at a site recommended by the Ministry of Science and ICT, staffed with resident engineers and a project manager ^[1][3][6]. The center's stated job is to develop and test an open, heterogeneous AI computing infrastructure that combines AMD CPUs and GPUs with domestically developed neural processing units, which places a non-Nvidia vendor at the center of a state effort to stop treating the GPU as the only unit of inference capacity ^[2][8].

The details were laid out at a forum on building an open AI computing infrastructure ecosystem, held in Jongno-gu, Seoul, on the 20th, presented as a follow-up to a memorandum of understanding that the Ministry of Science and ICT and AMD signed on the 23rd of the previous month ^[4][5]. Under that MOU the two sides agreed to jointly develop and test infrastructure connecting AMD CPUs and GPUs with domestic NPUs so the parts share AI computing tasks ^[7]. The reasoning given is demand-side: as inference demand grows rapidly, there is a need to move past GPU-centered infrastructure and combine various semiconductors, networks and software to optimize performance and power efficiency ^[8].

Read the commitments in sequence, because that is where the risk sits ^[1]. This year: the center, resident engineers, a PM ^[6]. By the end of next year: additional resident engineers and joint development and testing projects run with the ministry ^[9]. In 2027: Moon Kyung-sun, an executive director at AMD, said the company is reviewing a plan to deploy Helios, its latest integrated AI rack-scale system, at the center after preparations including a data center build are complete ^[10]. A review is not an order, and Helios is the only element that would involve substantial hardware.

The software leg is further along in the calendar. This year AMD and the ministry will test whether AMD's high-performance computing resources and its ROCm software are effective in advanced science and technology R&D, centered on the National AI for Science Center, with plans to apply them to national AI-for-Science projects from next year ^[11][12]. Moon said candidate fields include life sciences, climate and oceans, materials and semiconductors, space and astronomy, and large-scale simulations ^[13].

Governance arrives next month, when a steering committee launches with a ministry official and a senior vice president in charge of AI at AMD headquarters as co-chairs; the co-chairs plan a first online meeting in late September and regular meetings to develop projects and manage risks ^[14][15]. The ministry's ambition is to secure successful testing cases domestically and then turn them into a reference for a globally competitive open heterogeneous architecture, with the center acting as the point of contact between AMD, the ministry and domestic K-NPU developers ^[16][17]. Bae Kyung-hoon, Deputy Prime Minister and Minister of Science and ICT, said the government will strengthen full-stack testing support so it can quickly secure cases in which domestic NPUs are used at actual data centers and AI service sites ^[18].

Watch three things. Which K-NPU developers actually get silicon into the testbed, since the account of the forum names none, and reports no investment figure or purchase volume ^[2]. Whether the 2027 Helios review converts into a deployment. And whether this year's ROCm results are strong enough to justify the promised move into national AI-for-Science projects next year ^[11][12].