On July 28, according to Yonhap News Agency, South Korea’s Ministry of Science and ICT signed a memorandum of understanding with AMD on cooperation in the AI chip ecosystem. The two sides will establish an AI research center in South Korea, known as the “AI Center of Excellence,” with the aim of building an open AI infrastructure ecosystem. The two sides plan to develop heterogeneous AI computing infrastructure by connecting AMD’s CPUs and GPUs with AI chips, or NPUs, from South Korean companies that have strengths in inference. The cooperation will cover areas including memory, networking, servers, and software. The center will provide support for South Korean companies, universities, and research institutions in areas such as technical cooperation, AI chip software development, computing technology validation, and joint research. It will also promote collaboration with local South Korean AI chip companies.
According to semiconductor industry reports on the 20th, Samsung Electronics, SK Hynix and Micron have all stopped developing their own CXL (Compute Express Link) controllers, despite previously making major investments in this next-generation semiconductor technology. Micron was the first to terminate its proprietary controller development project and has shifted to PrimeMars’ solution. SK Hynix recently formally notified key partners that it would stop in-house development, with relevant R&D personnel reassigned to the processing-in-memory (PIM) field. Samsung Electronics will use its self-developed controllers only for internal R&D, while excluding them from commercial projects. For products sold externally, the company will instead procure controllers from fabless companies. The three manufacturers had originally planned to sell high-priced finished “integrated CXL modules” that combine self-developed controllers with DRAM. However, large data center customers, seeking to reduce infrastructure construction costs, prefer a “disaggregated” architecture: connecting CXL controllers separately to system motherboards while using standard DRAM DIMMs. The shift is driven by concerns over “self-cannibalization.” Forcing the adoption of expensive integrated products could weaken demand for standard DRAM DIMMs, which remain a core revenue source for memory manufacturers. The market gap created by this shift is expected to be filled by fabless companies such as Montage Technology, Astera Labs and PrimeMars.
On July 14, United Microelectronics Corporation (UMC) announced that its 12-inch silicon photonics production line in Singapore has begun volume production of advanced photonic chips, addressing growing demand for high-speed interconnects in AI data centers. UMC Senior Vice President Steven H. K. Hong said silicon photonics and CPO technologies will become important growth drivers for the company over the next several years. Shifting photonic chip production from traditional 8-inch wafers to 12-inch wafers can effectively reduce optical loss while improving power efficiency and transmission performance. Singapore-based photonic chip company Silith Technology has become the first core strategic customer of the production line. Its customers include optical module maker Zhongji Innolight and Coherent, both of which are major suppliers to Nvidia and Google. UMC is also jointly building a photonic chip production platform with Belgium-based microelectronics research center imec, with plans to open the platform to more customers starting in 2027. On the packaging side, UMC plans to launch customized advanced packaging services in 2027, connecting photonic chips to interposers to shorten electrical transmission paths. In its long-term roadmap, UMC plans to launch an open silicon photonics development platform in 2028 while also carrying out R&D on thin-film lithium niobate chiplet technology. China’s Taiwan region and Singapore will jointly serve as UMC’s main R&D and manufacturing bases for photonic chips and advanced packaging. The company has already assembled a specialized technical team of around 100 people.