“In the future, AI semiconductors will not be able to move to the Large Language Model (LLM) in all areas, and will focus on edge devices where differentiated data is important. And foundries with the technology to produce these semiconductors will play a key role in the AI semiconductor era.”

The Artificial Intelligence Semiconductor Forum held the '6th Artificial Intelligence Semiconductor Forum Breakfast Lecture' at the InterContinental Seoul COEX on the 26th. At this event, Intel Korea Vice President Lee Ju-seok presented on the topic of ‘Enabling a Trusted Design Ecosystem.’

Vice President Lee Ju-seok noted that the AI ecosystem should pay attention to whether it will open a semiconductor market with a specific purpose.

AI semiconductors are important for data, and whether it is CPU, GPU, or NPU, they are connected to the cloud, and as all devices are connected, data storage speed is important, and the network technology that follows is important, he said.

He said that the current data is being collected to comprehensively combine the data to obtain meaningful results at once and to derive beneficial results through correlations, analyze the data with AI to derive insights, and these insights play an important role in the AI market, but computing resources are needed to process this data, and the reason for creating AI semiconductors and NPUs is that there will continue to be demand in this data market.

In addition, since data is continuously generated, it is impossible to move to a Large Language Model (LLM) in all areas in the midst of a huge flood of data, and ultimately differentiated data processing will become important. He predicted that very small models such as healthcare, smart factories, smartphones & mobile devices, drones, agricultural sensors and equipment, smart home devices, wearables, IIoT, autonomous driving, and robots will process AI data on edge devices.

Accordingly, he argued that the market for various types of AI semiconductors in embedded form will grow explosively, and that the small-volume, multi-variety, and NPU markets will grow, and who will produce them will be the key.

The foundries that can make these AI semiconductors are currently TSMC, Samsung Electronics, and Intel, but this is because the technology is difficult and investment in foundries is difficult.He mentioned that this is because the amount of money involved is enormous and it takes at least 2-3 years.

In particular, he added that semiconductor lead times have become longer due to the past semiconductor supply and demand imbalance, and now semiconductor manufacturing companies also need to have the investment capacity to prepare for such situations.

Accordingly, Intel announced that it is formulating a strategy to prepare for such AI semiconductor future prospects and supply/demand imbalances, and is expanding foundries to prepare for technology and demand so that it can follow data trends such as CPUs, GPUs, and NPUs.

He also said that in AI inference, NPU power consumption accounts for 60-70% of total power consumption, so how to reduce power consumption at the edge is the most important technology, and that a lot of investment is being made in this area as well.

In addition, interface compatibility with CPU peripherals and how to implement software will be key factors in quickly transmitting data, and Intel has been continuously investing in software technology so that it can run on all hardware, from Gaudi to Xeon CPUs and through software called One API, and has also been continuously creating a developer community in Korea.