Don’t ignore the weight of HBM
LEE BYOUNG-HUNThe author is a professor of semiconductor engineering at Postech.
With the advent of ChatGPT, the era of artificial intelligence is now in full swing. The stock price of Nvidia, which monopolizes the market for graphics processing units (GPUs) — an essential tool in AI — has soared. The markets for the GPU board and the high-bandwidth memory (HBM) that go into GPUs are also expected to grow.
Even before reflecting the rapid growth trend, the HBM market was expected to grow to $2 billion in 2023 and $6.3 billion in 2028. The overall memory market is anticipated to grow to $136 billion from $110 billion during the same period. The share of HBM in the total memory market would grow from less than 2 percent to 4 percent.
The increase in AI services following the introduction of ChatGPT shows the potential for the HBM market to grow much faster than expected. As memory chipmakers show different degrees of preparation for HBM, companies that have fallen behind in technology development are rushing to catch up.
As HBM requires a stacking of multiple memory chips — and peripheral circuits to enable CPU/GPU and high-speed connection — a wide range of high-end technologies for foundry, memory, and chiplet integration packaging are needed. Because high-performance semiconductor systems such as AI chips mostly use HBM, it is important to enhance the fabless capability to design them.
In particular, the production technology used in memory driving circuits demands a cutting-edge, foundry-class technology that far exceeds the performance of conventional memory device drivers. That means Korea can have an advantage as it has already built both memory technology and high-end foundry technology.
HBM shows that it is meaningless to distinguish memory, foundry, and fabless (system integrated circuits, ICs) from now. If any one part is lacking, it will certainly deal a blow to overall competitiveness. One can only survive the fierce chip war by building the R&D ecosystem to secure unrivaled technologies against competitors, not to mention the cutting-edge ability to design and manufacture memory and system ICs by either enhancing domestic capabilities or establishing an international alliance.
As Japan’s alliance with the United States and Europe — and Taiwan and Japan’s active participation in U.S.-led chip technology development program — illustrates, competitors are moving fast to supplement their weaknesses in various fields of design, manufacturing, and packaging. But Korea does not have a national research organization representing the chip industry or a command center overseeing international cooperation.
Moreover, the government’s chip support strategy is primarily focused on training people and building a specialized semiconductor complex rather than developing super-gap technologies. The government’s meticulous effort to augmenting Korea’s existing strengths and supplement its apparent weaknesses are needed now more than ever.
with the Korea JoongAng Daily
To write comments, please log in to one of the accounts.
Standards Board Policy (0/250자)