Nam Hyun-woo has worked as a staff writer at The Korea Times since 2013, mostly covering business and politics. He currently belongs to the Business Desk where he covers topics such as emerging tech, AI, ICT and Korea's chaebol community. Prior to joining the team, he was the paper's correspondent for the presidential office of Korea during the Yoon Suk Yeol and Moon Jae-in administrations.
Chip guru introduces high-bandwidth flash road map for AI inference
Professor Kim Joung-ho of the Korea Advanced Institute of Science and Technology speaks during a press conference on high-bandwidth flash at the Korea Press Center in Seoul, Tuesday. Korea Times photo by Nam Hyun-woo
By Nam Hyun-woo
HBM pioneer Kim Joung-ho targets 2027-28 commercialization of HBF
Professor Kim Joung-ho of the Korea Advanced Institute of Science and Technology (KAIST), known for devising the basic structure of high-bandwidth memory (HBM), introduced a road map for high-bandwidth flash (HBF), a next-generation chip architecture for high-performance storage that will be critical for artificial intelligence (AI) inference.
“So far, graphics processing units (GPUs) have played a central role in AI computing, but going forward, memory will become even more important than GPUs,” Kim said during a press conference in Seoul on Tuesday.
“After years of research, HBM has been commercialized successfully and is now driving the AI era. With HBF, we aim to make another leap forward and lead the AI era through memory-centric computing.”
Memory, mostly dynamic random-access memory (DRAM) chips, serves as a fast, temporary and volatile workspace for active data and applications. Storage, largely based on NAND flash, functions as a non-volatile repository, retaining files and apps at larger capacity but slower speeds.
The importance of NAND flash has been rising rapidly in AI accelerators, as AI services increasingly shift toward inference, which requires fast and frequent access to large volumes of stored data to deliver real-time results.
While HBM is manufactured by vertically stacking DRAM chips, HBF is a new solution that vertically stacks NAND flash to maintain solid-state-drive-level capacity while aiming to achieve bandwidth expansion comparable to HBM. To enable this, technologies used in HBM production will be required.
A diagram showing a graphics processing unit next to high-bandwidth memory and high-bandwidth flash chips / Courtesy of KAIST Tera Lab
Kim proposed a structure in which HBF would be placed alongside the GPU and HBM, thereby allowing HBM to work as “a bookshelf near the GPU” and HBF “a library near the GPU.”
“In the PC era, Intel led the global market as CPU was the key component, while in the smartphone era, Qualcomm took the lead as power efficiency became critical,” he said. “In the AI era, memory will be the key determinant, with HBM driving speed and HBF defining capacity.”
Kim said HBF could be commercialized in late 2027 or early 2028, noting that design and process expertise accumulated through HBM development has significantly shortened the path to commercialization. He added that Samsung Electronics and SK hynix are working with Nvidia, Google, AMD and other major tech companies to roll out HBF-powered chips within that time frame.
“Companies need HBM capabilities to develop HBF as well,” he said. “Since Samsung Electronics and SK hynix possess technologies in both areas, their market dominance is expected to strengthen further compared to other memory companies. Going forward, it will become a hegemony game in which GPUs cannot operate without memory.”
Kim’s lab, KAIST Tera Lab, is known as a breeding ground for semiconductor experts. Around 40 Tera Lab alumni work at domestic memory giants including Samsung Electronics and SK hynix, while another 40 or so are employed at major U.S. tech firms such as Apple.
KAIST Tera Lab will hold an online Zoom presentation session on HBF technology development, workload analysis and its road map at 9 a.m. Feb. 10. A recording of the livestream will be available at the lab’s website (http://tera.kaist.ac.kr).