"Korean AI Semiconductor Technology Surpasses Nvidia...Urgent Need for Full Government Support" [Report]

Visiting AI Semiconductor Companies with the Minister of Industry
Visited Gaonchips, Libellion, Hanamicron, and Others
"Fabless, Foundry, and Packaging Ecosystem Must Be Well Established"

On the 17th, at the System Semiconductor Design Support Center located in Seongnam, Gyeonggi Province's Second Pangyo, Jeong Han-ul, CEO of Articron, which is based there, demonstrated a technology that uses a low-power artificial intelligence (AI) semiconductor currently under development to convert low-resolution images into high-resolution images in real time.

A snowy landscape photo with uneven image quality instantly transformed into a clear high-resolution image after passing through a PC board equipped with the AI semiconductor. Watching this, Minister Ahn Deok-geun of the Ministry of Trade, Industry and Energy and attending reporters asked numerous questions. Founded in 2022, this company is developing ultra-small low-power AI semiconductors. CEO Jeong explained, “Samsung Electronics gave us a significant discount for producing prototypes,” adding, “The System Semiconductor Design Support Center also covered about 70% of the costs.”

Articron CEO Jeong Han-ul is demonstrating technology that uses an ultra-small AI semiconductor under development to convert images in real time. Photo by Kang Hee-jong

On that day, about ten reporters covering the Ministry of Trade, Industry and Energy visited domestic system semiconductor companies and the Design Support Center in Second Pangyo along with Minister Ahn Deok-geun to observe the status of domestic AI semiconductor development and the ecosystem. Some of the domestic AI semiconductor companies visited appeared to have secured a considerable level of technological capability but still voiced difficulties and appealed for active government support.

"China requires 40% domestic chips for product shipment..."

The System Semiconductor Design Support Center located in Second Pangyo is a representative facility established to support fabless companies (semiconductor design-specialized companies) like Articron. Since its opening in 2020, it has supported 33 fabless startups. The center provides system semiconductor design companies with offices, design software, and measurement equipment at affordable prices. It has secured 103 copies of 33 types of Electronic Design Automation (EDA) tools, which cost tens of millions to hundreds of millions of won per copy.

Supporting prototype production, which fabless companies cite as their biggest challenge, is also one of the center’s main tasks. For multi-project wafers (MPW) using a 10-nanometer (nm) process, support can reach up to 560 million won. However, the secured budget is only about 6 billion won annually, which is far from sufficient.

Yu Byung-du, head of fabless support at the Korea Semiconductor Industry Association, explained, “Typically, producing prototypes with an 8nm process costs 20 billion won, and 5nm costs 30 to 40 billion won,” adding, “Since 3 to 4 prototype productions are needed for commercial testing, 100 billion won is required.” This amount is daunting for fabless startups to produce prototypes using fine processes.

Yoo Byung-du, Head of Fabless Support Office at the Korea Semiconductor Industry Association, is explaining the System Semiconductor Design Support Center. Photo by Ministry of Trade, Industry and Energy

Gaonchips, located about a five-minute walk from the Design Support Center, is a semiconductor design house and an official design solution partner of Samsung Electronics Foundry. A design house acts as a bridge connecting fabless and foundry (semiconductor contract manufacturing) companies. It provides services that convert design blueprints created by fabless companies into formats suitable for foundry production processes.

In the past, fabless companies directly requested product manufacturing from foundries, but as processes have become finer, such as 3nm and 2nm, the help of design houses like Gaonchips has become necessary. Especially recently, with the rise of AI semiconductors, related sales have increased significantly. The average annual growth rate from 2020 to this year exceeds 50%. AI and automotive semiconductor sales account for more than 70% of total sales.

CEO Jeong Gyudong of Gaon Chips is explaining the company's status to Minister Ahn Deokgeun of the Ministry of Trade, Industry and Energy. Photo by Ministry of Trade, Industry and Energy

However, domestic design houses are at a disadvantage compared to Taiwan, a major competitor of South Korea. It is known that GUC, a subsidiary of TSMC and a design house, has annual sales exceeding 1 trillion won.

Jeong Kyu-dong, CEO of Gaonchips, emphasized that for the domestic system semiconductor industry, including AI semiconductors, to grow, the government needs to provide active incentives so that companies can use more domestically produced semiconductors. CEO Jeong said, “China has a system where products can only be shipped if at least 40% of the semiconductors used are domestically produced,” adding, “South Korea should also have systems such as tax benefits if more domestic semiconductors are used.”

"NVIDIA and Qualcomm employees also joined... shortage of semiconductor architects"

Next, Rebellion, an AI semiconductor company that recently attracted attention by raising 165 billion won in Series B funding, was visited. Established in 2020, it developed two AI semiconductors in just over two years and has raised a cumulative investment of 280 billion won in about three years. Rebellion is developing a neural processing unit (NPU) to replace NVIDIA’s graphics processing units (GPUs). Former developers from NVIDIA, Google, and Qualcomm have also joined Rebellion.

On that day, Jinwook Oh, Rebellion’s co-founder and Chief Technology Officer (CTO), demonstrated a comparison between the company’s ATOM NPU and NVIDIA’s A100 GPU to reporters. They compared the power consumption when generating images using generative AI or responding to questions with a large language model (LLM) on servers equipped with the two chips.

Jinwook Oh, Chief Technology Officer (CTO) of Rebellion, is demonstrating a comparison between the company's Atom NPU and NVIDIA A100 GPU to reporters who visited on the 17th. Photo by Heejong Kang

The results showed that the ATOM NPU consumed a maximum of 92 watts (W) during image generation, while the A100 consumed 326W. Power consumption during language generation was 74W and 281W, respectively. CTO Oh explained, “Energy efficiency is 3 to 4 times better compared to the latest GPU technology.” Rebellion’s NPU is an AI-optimized semiconductor that offers superior performance to NVIDIA’s GPU while being supplied at about 30% lower cost.

Currently, Rebellion is developing the next-generation AI semiconductor ‘REBEL’ targeting large-scale AI together with Samsung Electronics. This semiconductor, manufactured using a 4nm process, will be equipped with HBM3E, a high-performance memory for AI. It supports multimodal processing of various data types such as text, images, voice, and video, and is scheduled for release in the second half of 2024.

Rebellion was founded by CEO Park Sung-hyun, a former Morgan Stanley executive in the U.S., and CTO Oh, a former IBM developer. CTO Oh cited excellent talent, ease of funding, and government support as reasons for starting the company in South Korea rather than the U.S. He explained, “There are many talented people trained at good companies in Korea, and government support is available for about three years during the early startup phase, making it a good environment for latecomers to start a business.” Rebellion also received government support by moving into the System Semiconductor Design Support Center.

However, CTO Oh noted, “There are many semiconductor designers in Korea, but the number of architect personnel who design the overall structure is insufficient compared to overseas,” adding, “Gaining extensive operational experience is necessary to narrow the gap with advanced countries.”

"Taiwan freely collaborates centered on TSMC... we even struggle to obtain products"

In the afternoon, the headquarters of Hana Micron in Eumbong, Asan, Chungnam, was visited. Hana Micron is a semiconductor back-end process specialist that receives wafer-form semiconductors from Samsung Electronics, SK Hynix, and others, then thins, cuts, attaches them to substrates, and molds them.

Hana Micron focuses on packaging businesses for system semiconductors and memory semiconductors used in smartphones and PCs. Recently, it has been developing advanced packaging technology that integrates NPUs and high-bandwidth memory (HBM). The technology development is expected to be completed by the end of this year.

Hana Micron is the number one company domestically in the semiconductor back-end process field but ranks 11th globally. The top company is Taiwan’s ASE, followed by the U.S.’s Amkor. Among the top 10 worldwide, six companies are Taiwanese.

Core back-end process equipment such as die bonding still heavily depends on Japan. Park Jin-ho, head of packaging technology at Hana Micron, said, “Domestic equipment accounts for about 30 to 40%,” adding, “The level of domestic equipment is improving, and it is expected to expand to 50%.”

Semiconductor companies suggested that for the domestic AI semiconductor industry to thrive, a well-established ecosystem connecting fabless, foundry, and back-end processes is essential. Lee Dong-chul, CEO of Hana Micron, said, “In Taiwan, companies freely collaborate centered on TSMC, but we even struggled to obtain products for developing AI semiconductor packaging technology,” adding, “It would be good if supply chain management (SCM) is well established in the future.”