Professors discuss the memory revolution in the AI era—The transition from HBM to HBF

Recently, Professor Jung-Ho Kim, a leading researcher in semiconductor memory at the Korea Institute of Science and Technology, provided important insights into the major shift in memory technology driven by the development of AI. According to his comments via email responses, the enhancement of artificial intelligence’s thinking and reasoning capabilities is creating challenges that cannot be addressed by traditional memory configurations. Professor Kim pointed out at an industry briefing on “HBF Research Content and Technology Development Strategy” that shifting from text to voice interfaces is an unavoidable reality, and he issued a warning about the rapid increase in data requirements associated with this transition.

AI Evolution and Structural Changes in Memory Demand

In conventional systems, up to two graphics processing units (GPUs) are connected vertically, operating alongside high-bandwidth memory (HBM). However, it has become clear that this configuration cannot meet the data processing needs required for advanced AI functions such as speech recognition and text generation. Professor Kim notes that at the current growth rate of HBM, it will become physically difficult to cope with the surging AI demand, and the industry will be forced to transition to HBF.

Coexistence of HBM and HBF and the New Memory Architecture

As a future memory strategy, simultaneous deployment of HBM and HBF is being considered to overcome capacity limitations. Furthermore, in more advanced stages, the development of MCC (Memory-Centric Computing) architecture, which organically integrates CPU, GPU, and memory on a single core chip, could dramatically improve overall system efficiency. From his email responses and speeches, it is clear that such structural shifts are an inevitable response to the demands of the AI era.

The Era When HBF Surpasses HBM in 2038

Professor Kim’s analysis predicts that after 2038, the demand for HBF will surpass that for HBM. This timeline suggests how rapidly the continuous evolution of AI technology, exponential increases in data processing, and fundamental changes in memory system design will be realized. His email responses and industry briefings repeatedly emphasize the importance for companies and research institutions to prepare now for this transition. It is evident that this is not merely a technological trend but an unavoidable industrial transformation that must be recognized.