[Chip Talk] As AI Adoption Increases, Samsung and SK Gain Advantage... Critical Race in Next-Generation Semiconductor Development

HBM-PIM, 3D DRAM Semiconductor Development
Future Perceptron AI Semiconductor Competition

"Big Tech Companies Likely Won't Make Memory"

[Asia Economy Reporter Moon Chaeseok] Memory semiconductor companies such as Samsung Electronics, SK Hynix, and the US-based Micron are fiercely competing in the development of artificial intelligence (AI) semiconductors. This is because companies like Microsoft and Google have launched conversational AI service programs. The speed of developing advanced semiconductors for AI services is also expected to increase. Memory semiconductor companies have planned to create high-bandwidth intelligent memory (HBM-PIM) semiconductors and next-generation 3D (three-dimensional) DRAM semiconductors to be integrated into AI programs.

The industry has long anticipated that DRAM demand will surge as AI products increase. This is because efficient DRAM is needed to quickly store calculation results in AI computers. AI semiconductors perform many data calculations simultaneously using parallel processing methods. To enhance AI performance, the parallel data calculation time of deep learning algorithms must be reduced. For smartphone or cloud computer AI services, GPUs (graphics processing units) and DRAM must be able to exchange data within 1 millisecond (one-thousandth of a second).

OpenAI logo and ChatGPT. [Photo by AFP Yonhap News]

The most effective AI semiconductor parallel processing device is the GPU. GPUs incorporate HBM (high-bandwidth memory) semiconductors. SK Hynix supplies HBM3 semiconductors to Nvidia, the world’s leading GPU company based in the US. Samsung Electronics is developing HBM-PIM semiconductors, which combine HBM with intelligent memory (PIM) semiconductors, together with AMD, the world’s second-largest GPU company in the US. HBM-PIM semiconductors are components that combine AI processors with computational capabilities into memory semiconductors primarily designed for data storage.

Following HBM-PIM semiconductors, the next focus is on 3D DRAM. It is expected that 3D DRAM technology development will be completed in about three to four years. Memory semiconductor companies are releasing 3D DRAM IP (intellectual property). Currently, products that improve efficiency by reducing DRAM cell size are used for 3D processing. Once 3D DRAM becomes widespread, competition will shift to changing the orientation or shape of the cells.

As AI services develop, the era of high-capacity server DRAM is expected to arrive sooner. Currently, 64GB server DRAM is mainly supplied. Transitioning to 128GB is only a matter of time. With the release of AI service programs, speeds are expected to increase further. Samsung Electronics and SK Hynix stated during their Q4 earnings conference calls last year that "the server memory semiconductor market will accelerate the transition from 64GB-centered to 128GB-class."

Will companies like Microsoft and Google make AI semiconductors themselves? Experts say "the likelihood is low." The analysis is that the practical benefits are not significant. AI programs developed by big tech companies require higher computational capabilities rather than storage functions. Therefore, there is no need to develop memory semiconductors themselves.

As the AI semiconductor market grows, memory semiconductor companies are expected to gain an advantage. Kim Yangpaeng, a senior researcher at the Korea Institute for Industrial Economics & Trade, said, "Memory semiconductors are mainly limited to storage functions, so big tech companies are unlikely to develop their own memory chips." Hwang Cheolsung, a distinguished professor in the Department of Materials Science and Engineering at Seoul National University, also said, "Unless big tech companies manufacture DRAM, the increase in AI services will benefit DRAM manufacturers."

Attention should also be paid to when and which company will develop new low-power AI semiconductors after HBM-PIM and 3D DRAM. To create low-power AI semiconductors, the perceptron concept must be implemented within the internal circuitry. The perceptron is an artificial neural network model used in deep learning. Once perceptron semiconductors become widespread, it will be possible to embed high-density processors at the AI computer level within a semiconductor. Kim Jeongho, director of the KAIST Global Strategy Institute, said, "When new low-power AI semiconductors implementing the perceptron concept become widespread, AI functions will be embedded in all objects," adding, "Samsung Electronics, SK Hynix, and US companies such as Intel, IBM, AMD, and Nvidia will compete to capture this AI processor market."

This content was produced with the assistance of AI translation services.