AI NewsWaveAMD Teams With FuriosaAI and MangoBoost on Heterogeneous AI Infrastructure in Seoul
Key points
- AMD signs heterogeneous AI infrastructure agreement with two Korean firms
- Plan to build demonstration infrastructure in Seoul next spring
- Ministry of Science and ICT starts demonstration combining domestic NPUs
AMD said at an event in Seoul on the 4th that it will work with Korean AI chip companies FuriosaAI and MangoBoost to build open heterogeneous AI infrastructure that links different types of processors. AI Newspaper (인공지능신문), Digital Daily, IT Chosun, iNews24, Bloter, The Elec and AI Times reported the same news. Digital Daily, Bloter and AI Times wrote that the cooperation was formalized as a memorandum of understanding (MOU).
Descriptions of the setup differed slightly by outlet. Bloter wrote that it combines AMD's central processing units (CPUs) and graphics processing units (GPUs) with Korean companies' neural processing units (NPUs) and data processing units (DPUs). AI Newspaper (인공지능신문) also reported that the partners will develop computing infrastructure suited to AI inference by combining AMD's CPU and GPU technology with Korean NPU and DPU technology. IT Chosun described it as open infrastructure combining GPUs, NPUs and DPUs. Digital Daily introduced MangoBoost as a DPU specialist, while Bloter described it as a data center infrastructure company.
Expressions about timing and scale also varied. iNews24 reported that the companies plan to build AI infrastructure in Seoul next spring and verify performance and efficiency by linking different chips into one system. AI Times wrote that a demonstration infrastructure for a heterogeneous AI cluster will be built in Seoul next spring. The Elec reported that the companies will jointly build a heterogeneous AI computing infrastructure cluster in Seoul next year, adding that Keith Strier, AMD's senior vice president for global AI markets, said so at the event held on the 4th at the Westin Seoul Parnas in Samseong-dong. The event name was written differently: The Elec called it "AMD Advancing AI Seoul 2026," while AI Times called it "AMD Advancing AI World Tour Seoul 2026."
The government and Korean companies also made moves at the same event. Digital Daily reported that the Ministry of Science and ICT will launch a large-scale demonstration project with AMD that combines domestic NPUs, in order to break the monopoly of specific foreign GPUs and secure national AI infrastructure sovereignty. The same outlet conveyed in a headline that Vice Minister Ryu Je-myeong of the ministry said he would use a successful case of combining AMD and domestic NPUs next year. Moreh unveiled "Token Factory" software that lowers AI inference costs on AMD GPUs. According to iNews24, Moreh said it secured more than twice the AI inference throughput of Nvidia's B200 using AMD's Instinct MI355X. This figure is what Moreh stated.
Recent developments Lisa Su, AMD's chief executive officer (CEO), said at the third-quarter earnings call on the 3rd (local time) that she expects next year's annual revenue to grow more than 80% from this year, according to The Elec. Server CPU revenue is expected to double, and AI GPU revenue is expected to grow faster than that. Chosun Biz reported that AMD posted record third-quarter results on rising demand for AI data center chips, with operating profit up 123% from a year earlier. iNews24 reported that Lisa Su called Korea a key partner in the AI industry and said AMD will expand investment and cooperation.
Terms explained: Heterogeneous computing is a method of using processors with different characteristics, such as CPUs, GPUs and NPUs, together in one system. An NPU is a chip specialized for AI computation, and a DPU is a chip that handles data movement and processing between the network and storage devices.