KAIST develops human brain-like AI semiconductor
Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed an AI semiconductor capable of processing large language model (LLM) data at ultra-high speeds while significantly reducing power consumption, according to the Ministry of Science and ICT.
The ICT ministry said Wednesday that a research team led by professor Yoo Hoi-jun at KAIST’s processing-in-memory research center developed the world’s first complementary-transformer AI chip using Samsung Electronics’ 28-nanometer manufacturing process.
The complementary-transformer AI chip is a neuromorphic computing system that mimics the structure and function of the human brain. Utilizing a deep learning model commonly used in visual data processing, the research team successfully implemented this transformer function, gaining insights into how neurons process information.
This technology, which learns context and meaning by tracking relationships within data, such as words in a sentence, is a source technology for generative AI services like
ChatGPT, the ministry said.
The research team demonstrated the functionality of the complementary-transformer AI chip at the ICT ministry’s headquarters in Sejong on Wednesday.
Kim Sang-yeob, a member of the research team, conducted various tasks such as sentence summarization, translation and question-and-answer tasks using OpenAI’s LLM, GPT-2, on a laptop equipped with a built-in complementary-transformer AI chip, all without requiring an internet connection. As a result, the performance
was notably enhanced, with the tasks completed at least three times faster, and in some cases up to nine times faster, compared to running GPT-2 on an internet-connected laptop.
To implement LLMs typically utilized in generative AI tasks, a substantial number of graphic processing units (GPUs) and 250 watts of power are typically required. However, the KAIST research team managed to implement the language model using a compact AI chip measuring just 4.5 millimeters by 4.5 millimeters.
“Neuromorphic computing is a technology that even companies like IBM and Intel have not been able to implement, and we are proud to be the first in the world to run the LLM with a low-power neuromorphic accelerator,” Yoo said.
He predicted this technology could emerge as a core component for on-device AI, facilitating AI functions to be executed within a device even without requiring an internet connection. Due to its capacity to process information within devices, on-device AI offers faster operating speed and lower power consumption compared to cloud-based AI services that rely on network connectivity.
“Recently, with the emergence of generative AI services like ChatGPT and the need for on-device AI, demand and performance requirements for AI chips are rapidly increasing. Our main goal is to develop innovative AI semiconductor solutions that meet these changing market needs. In particular, we aim to focus on research that identifies and provides solutions to additional problems that may arise during the commercialization process,” Yoo added.
The research team said this semiconductor uses only 1/625 of the power and is only 1/41 the size of Nvidia’s GPU for the same tasks.