AI Semiconductors
Samsung Electronics’ Foundry Business Division has started volume production of next-generation artificial intelligence (AI) chips in earnest, starting with foundry orders from Korean fables companies. Thanks to the growing AI chatbot craze fueled by the popularity of ChatGPT, the AI semiconductor market is on a steep rise, the Korean semiconductor giant is expected to land more foundry orders from large customers.
According to industry sources on April 5, Samsung Electronics’ Foundry Business Division recently started mass-producing the first-generation chip called “Warboy,” a neural network processing unit (NPU), after taking an order from AI-specialized fabless FuriosaAI in a 14-nm process. The second-generation Warboy chip with improved performance will roll out via Samsung Electronics’ 5-nm process line during the first half of next year. The Atom of Rebellions Inc., a Korean AI chip startup that specializes in designing AI semiconductors, was also mass-produced at Samsung Electronics’ 5-nm foundry process line.
Current AI platforms are primarily equipped with graphics processing units (GPUs) that can simultaneously perform many computations in parallel. More than 10,000 high-performance Nvidia A100 GPUs for data centers were used to realize ChatGPT services.
NPUs were designed exclusively for AI to compensate for the weaknesses of GPUs such as high power consumption and high prices. Their processing speed is 10 times faster than that of GPUs. NPUs’ price and power consumption are only one-fifth to one-sixth of those of GPUs. As the creation of AI ecosystems is accelerating due to the ChatGPT craze, foundry orders for AI semiconductors from fables companies are on the rise.
Large customers are turning their eyes first to TSMC, the No. 1 company in the global foundry industry. Nvidia, the world’s No. 1 GPU powerhouse in the AI semiconductor market, is one of TSMC’s major customers. TSMC’s 7-nm and 4-nm process lines are making Nvidia A100 and H100 chips. AMD is also scaling up GPU orders to the Taiwanese foundry giant. When a foundry plant currently being built by TSMC in the U.S. state of Arizona is completed, orders from major customers such as Nvidia are expected to further increase. TSMC plans to start volume production of 4-nm chips in 2024 and 3-nm chips in 2026, respectively.
Samsung Electronics is receiving foundry orders from Korean fabless companies but the scales themselves are small. However, as global companies are scrambling to advance into AI semiconductors, driving the growth of the foundry market. This gives the Korean semiconductor giant much anticipation for landing big orders. Some industry insiders are raising the possibility that Nvidia will give a foundry order to Samsung Electronics for smooth chip supply. Google plans to upgrade and re-launch AI chatbot Bard which was introduced as Google’s answer to ChatGPT. Baidu, China’s largest e-commerce company, launched Ernie Bot, a Chinese version of ChatGPT. Microsoft announced that it will introduce a new service that incorporates ChatGPT into its search engine Bing. Naver will release Search GPT, a combination of AI services and its search services during the first half of this year.
Samsung Electronics is also focusing on raising its AI semiconductor production capacities by significantly ramping up facility investment. Production Line P4 is being built on the Pyeongtaek campus in Korea with the goal of starting operation early next year. It will spend 300 trillion won (US$228 billion) to build the world’s largest semiconductor cluster in Yongin, Gyeonggi Province of Korea by 2042. The Taylor City plant in Texas in the United States will mass-produce 4-nm chips. It is under construction and will go into operation in 2024.