"Our goal is to successfully develop AI semiconductor chips and complete proof-of-concept (PoC) projects with domestic and overseas customers."
Kim Joo-young, CEO of HyperAccel, told The Seoul Economic Daily at his office in Gangnam, Seoul, recently. HyperAccel, a startup founded in 2023, is developing an AI semiconductor chip called the LPU, specialized for large language models (LLMs). The LPU is a type of neural processing unit (NPU) that maximizes memory bandwidth utilization efficiency and applies LPDDR-based low-power memory to reduce power consumption. Through this approach, HyperAccel is implementing chips at one-third the power and one-tenth the cost of graphics processing units (GPUs), which are primarily used for LLM development.
"GPUs don't have high sustainability when it comes to power issues," Kim said. "When designing a chip, you have to decide what external memory to use, and HyperAccel chose LPDDR-based low-power memory through a specialized technical structure."
HyperAccel is currently developing both data center chips and on-device chips. The data center chip was designed in 2024 and produced in March this year. The on-device version is being developed by reducing the size of the data center chip to one-tenth. "We are co-developing with companies that could become potential customers to actually purchase the chips," Kim said. "Data centers can use electricity freely, but home appliance robots have power limitations, making on-device versions more complex." He added, "Wearable devices and mobile must use under 1 watt, home appliance robots under 10 watts, and vehicles under 100 watts. To meet power standards while also satisfying unit price requirements, you have no choice but to use specialized chips."
HyperAccel is mass-producing its data center chips at Samsung Electronics (005930.KS) and its on-device chips at TSMC. "This year we will produce hundreds of data center chip samples for PoC," Kim explained. "After that, samples of the on-device chip will be available by the end of this year." Currently, Naver Cloud and LG Electronics (066570.KS) are cited as key PoC customers. For Naver Cloud, the chips are expected to be used to run HyperCLOVA services around September. With LG Electronics, HyperAccel is verifying the performance of its on-device chips in humanoid robots and robot vacuum cleaners.
HyperAccel's focus on LPU chip development comes as the AI inference market continues to grow. According to global market research firm MarketsandMarkets, the AI inference market is projected to reach $254.9 billion by 2030, growing at an average annual rate of 19% from last year. As the AI inference market expands, demand for chips that can serve as complements to GPUs, which carry high cost and power burdens, will inevitably grow. "Of Nvidia's market capitalization of 5,000 trillion won, even securing just 1% for NPUs would be 50 trillion won ($36.9 billion)," Kim said. "Big Tech companies like Nvidia and Google are also focusing on AI-specific chips, so competition will intensify in earnest."
The fact that only a handful of countries—including Korea, the United States, China, Europe, Israel, and Taiwan—can develop NPUs is a favorable factor for Korea. This is because few countries can both effectively utilize AI and possess the hardware and software technical capabilities. The industry is also paying attention to the fact that the U.S. government is restricting GPU exports to China and Russia. "China is also showing great interest in NPUs," Kim said. "Government support is needed so that Korean companies can sell chips not only in the U.S. but also in Southeast Asia."
