From a Minor Player in System Semiconductors to a Leading Nation in Next-Generation AI Semiconductors

Sapeon Unveils New NPU 'X330'
FuriosaAI, Rebellion, and DeepX Also Active

Enhancing NPU Technology and Expanding Market Supply
Industry Revenue Growth Expected Next Year

Semiconductors are divided into memory semiconductors (DRAM, flash) that store information and system semiconductors (CPU, GPU) that calculate and process information. South Korea ranks first in the world in the memory semiconductor sector. Samsung Electronics and SK Hynix are the world's top two companies in this field. However, when it comes to system semiconductors, South Korea was not visible. The U.S. company Intel has virtually monopolized the computer CPU (Central Processing Unit) market, a representative system semiconductor, for over 40 years. The GPU (Graphic Processing Unit) market for graphic video system semiconductors has been dominated solely by the U.S. company Nvidia. Nvidia GPUs have recently gained more fame as AI computing devices rather than their original purpose of graphic computation.

New NPU 'X330' Unveiled by Sapeon /
[Photo by Sapeon]

However, as AI begins to be used more widely, there are forecasts that Korean companies will exert influence in the system semiconductor market. On the 16th, Sapeon, an AI semiconductor company spun off from SK Telecom, unveiled the 'X330,' an AI inference NPU. NPU is considered the next-generation AI semiconductor. It is a system semiconductor optimized for AI tasks such as deep learning. Compared to GPUs currently used as AI system semiconductors, NPUs consume less power and deliver higher performance relative to input. The semiconductor industry currently uses GPUs or CPUs as AI semiconductors, but it is expected that as AI technology advances, dedicated processors like NPUs will be primarily used.

The X330 is a product that improves computing performance by four times and power efficiency by two times compared to Sapeon's previous NPU model, the X220. It also outperforms Nvidia's GPU of the same class released this year by about twice in computing performance and 1.3 times in power efficiency. It is optimized for implementing large language models (LLM), which have risen alongside generative AI.

Recently, many Korean fabless (semiconductor design) startups have been competing in the NPU development field. Companies such as Sapeon, FuriosaAI, Rebellions, and DeepX have gained recognition for their technology through their self-developed NPUs. In fact, Rebellions recorded higher performance than Nvidia and Qualcomm with its NPU 'Atom' at MLPerf, a competition showcasing AI semiconductor technology, held last April.

This technological competitiveness is leading to business expansion. FuriosaAI has supplied its NPU 'Warboy' to cloud companies like Kakao Enterprise and Naver Cloud, as well as various enterprise customers. Rebellions supplied NPU prototypes to IBM in the U.S. and is accelerating business cooperation with KT, which recently invested 30 billion KRW in the company. KT has announced that it will use Rebellions' NPU in the process of expanding the application of its self-developed super-large AI 'Mideum.'

Kim Hyung-jun, head of the Next-Generation Intelligent Semiconductor Business Group, said, "Nvidia GPUs are general-purpose, consume a lot of power, and above all, there is a shortage of supply, so there is a high demand for alternatives." He added, "If domestic companies introduce NPUs that consume less power, there will definitely be market competitiveness." He also said, "The government is trying to expand the adoption of domestic NPUs in domestic data centers," and "If companies secure domestic references, it will greatly help them expand their presence in the global market."

AI semiconductor companies are showing confidence in expanding their performance by consecutively launching follow-up NPU models. Ryu Su-jeong, CEO of Sapeon, said, "We achieved meaningful sales through the X220 this year," and predicted, "Sales of the X330 will begin as early as the end of this year or the first half of next year." Kim Jung-wook, vice president of DeepX, also said, "We are about to start full-scale product commercialization," and "Next year, we will be able to share business results and achieve significant performance."

Meanwhile, the activities of these companies are expected to benefit domestic memory companies such as Samsung Electronics and SK Hynix. High-performance memory that supports processors like NPUs is essential for their operation. In fact, high-bandwidth memory (HBM), which has seen a surge in demand when used with GPUs, is also used in large quantities in NPUs. Not only Sapeon but also many companies such as Rebellions and FuriosaAI have announced the development and mass production of NPUs applying the latest HBM models.