The Future of AI Memory: Transitioning Power from GPU to High-Bandwidth Flash
Summary:
- Kim Jong-ho, the “father of HBM,” forecasts a pivotal shift in AI technology from GPU dominance to memory-centric configurations.
- High-Bandwidth Flash (HBF) is expected to evolve significantly by 2026, emerging as a key player in AI systems.
- Industry leaders, including SK Hynix and Samsung, are actively developing HBF technologies, signaling a new era in memory architecture.
The landscape of artificial intelligence (AI) technology is rapidly evolving, and a transformative shift is underway. Recent insights from Kim Jong-ho, a prominent figure in memory technology and a professor at the Korea Advanced Institute of Science and Technology (KAIST), reveal that the control of power in the AI era is pivoting from Graphics Processing Units (GPUs) to memory solutions. This transition is underpinned by emerging innovations in High-Bandwidth Flash (HBF), a next-generation memory technology that is set to play a crucial role in future AI applications.
The Emergence of High-Bandwidth Flash
In a pivotal statement on his YouTube channel, Kim underlined the growing significance of memory in the AI domain, predicting that HBF will become the new battleground following the establishment of High Bandwidth Memory (HBM). He anticipates that HBF will make notable advancements by early 2026, with its official entry into the market expected around 2027-2028. This forecast is grounded in the understanding that memory’s function is becoming increasingly critical as the demands of AI applications intensify.
Advantages of High-Bandwidth Flash
Conceptually, HBF shares similarities with existing HBM technology; both make use of through-silicon vias (TSVs) to stack multiple chip layers vertically. However, HBF is built on the foundation of NAND flash memory, offering substantial advantages in both capacity and cost efficiency. While NAND typically operates at slower speeds compared to Dynamic Random Access Memory (DRAM), its ability to expand in capacity by a factor of ten makes it a compelling alternative for modern AI solutions. By employing advanced layering techniques, HBF is poised to meet the burgeoning storage needs of AI models.
Industry Collaborations and Innovations
As this memory revolution unfolds, industry leaders are not standing idly by. Notably, SanDisk and SK Hynix entered into a memorandum of understanding in August 2025 to collaborate on the development and standardization of HBF technical specifications. The first HBF memory samples are slated for release in the latter half of 2026, signaling an impending wave of AI inference systems leveraging this new technology.
Additionally, at the OCP Global Summit in October 2025, SK Hynix showcased its pioneering AIN Family of storage products, featuring HBF technology. Similarly, Samsung Electronics has embarked on an early concept design process for its own HBF offerings, leveraging its extensive experience in high-performance storage solutions. Although specifics on mass production remain undisclosed, the initiative underscores the urgency among tech giants to embrace this evolving memory landscape.
Memory Architecture: A Multi-Level System
In elucidating the future of AI memory, Kim likens the memory hierarchy to a smart library:
- SRAM within GPUs functions as a desktop notebook, prioritizing speed and size.
- HBM serves as a side bookshelf, facilitating rapid access and calculations.
- HBF operates as an underground library, storing vast amounts of AI knowledge and continuously supplying data to HBM.
This metaphorical framework illustrates how these technologies will coalesce, allowing GPUs to integrate both HBM and HBF in complementary configurations. This integration will signify a new era where computing and storage capabilities are intricately linked, paving the way for unparalleled advancements in AI performance and efficiency.
Conclusion: A New Frontier for AI Technology
As the tech industry braces for the advancements brought forth by HBF, the implications for AI performance are profound. The transition of power from GPU-centric architectures to memory-oriented solutions like HBF will redefine how we approach AI applications and data processing challenges. By harmonizing storage and computing, we stand on the cusp of a new frontier in artificial intelligence, marking the beginning of an era characterized by enhanced capabilities and unprecedented efficiency.
In summary, as Kim Jong-ho aptly points out, the future of AI will heavily rely on innovative memory technologies like HBF, transforming how we interact with and leverage AI systems in our daily lives. This evolution signals that memory is not just a supporting player but a leading force in the advancement of artificial intelligence, shaping the technological landscape for years to come.
