김주선 SK하이닉스 사장(AI Infra 담당)이 지난 4일 SEMICON TAIWAN에서 ‘AI 메모리 기술의 새로운 가능성을 열다(Unleashing the possibilities of AI memory technology)’를 주제로 키노트를 진행했다.
HBM4 Development of just-in-time supply to meet customer demand is progressing smoothly
Mass production of GDDR7 supporting up to 40Gbps is ready
Kim Joo-sun, President of SK Hynix (in charge of AI Infra), announced that SK Hynix will begin mass production of HBM3E 12-layer products starting at the end of September and is developing HBM4 smoothly so that it can be supplied in a timely manner according to customer demands.
Kim Joo-sun, President of SK Hynix, gave a keynote speech titled ‘Unleashing the possibilities of AI memory technology’ at SEMICON TAIWAN on the 4th.
This presentation was made at SEMICON TAIWAN 'CEO Summit'.
CEO Kim Joo-sun said that in order for AI to develop and reach the level of AGI (Artificial General Intelligence), it is necessary to solve difficult problems related to power, heat dissipation, and memory bandwidth.
One of the big issues is power, with data centers expected to use at least twice as much power as they do today by 2028, and new forms of energy such as small modular nuclear power plants may be needed to provide sufficient power, he said.
In addition, as more power is used in data centers, the heat generated proportionally increases, so in order to continue to develop AI technology, effective solutions to solve the heat problem must be found, and for this purpose, SK Hynixannounced that it is working with partners to develop high-efficiency AI memory that can reduce heat generation by minimizing power usage while providing high capacity and high performance.
He also said that the demand for ultra-high-performance memory suitable for AI implementation is increasing, and that while bandwidth-related issues were not that important before ChatGPT was introduced, the demand for improved memory bandwidth is growing as AI technology advances.
To overcome these obstacles, SK Hynix stated that it is currently supplying HBM3E, high-capacity server DIMM, QLC-based high-capacity eSSD, and LPDDR5T to the market.
SK Hynix has been supplying HBM3E 8-layer products for the first time in the industry since the beginning of this year, and plans to begin mass production of HBM3E 12-layer products at the end of this month.
Compared to general servers, AI servers require more than four times the memory capacity, and to meet this demand, the company is supplying 256GB DIMMs for TSV technology-based servers.
SK Hynix is also the only supplier mass-producing QLC-based high-capacity eSSD, and plans to introduce a 120TB model that will significantly contribute to power efficiency and space optimization in the future. Finally, LPDDR5T is a product optimized for on-device AI that processes data at a speed of 9.6 gigabits per second.
The development of products and technologies for the future is also progressing smoothly. SK Hynix is developing HBM4 smoothly so that it can supply it in a timely manner according to customer demand. HBM4, which is the first to apply logic technology to the base die, will be produced through collaboration with TSMC and will demonstrate the best performance. In addition, it is steadily preparing next-generation memory products such as LPCAMM, CXL, and 512GB high-capacity DIMM.
In the NAND field, SK Hynix will also continue to develop cutting-edge products. Additionally, SK Hynix is nearing the final stages of mass production of the industry’s highest-performance GDDR7 supporting up to 40 Gbps, and is also developing LPDDR6 with innovative bandwidth and power.
President Kim Joo-sun said that in addition to R&D investment for product and technology development, infrastructure investment will proceed as planned, and that SK Hynix will build a state-of-the-art production facility in the Yongin Semiconductor Cluster, where land development construction is progressing smoothly, and that based there, it will engage in close cooperation with various global partners.
SK Hynix also said it plans to build a state-of-the-art packaging plant and R&D facility in Indiana, USA, with the goal of mass production in 2028, which will help strengthen collaboration with key customers and partners.