By Lexi Bryant
KAIST Unveils Revolutionary AI Semiconductor
A research team at the Korea Advanced Institute of Science and Technology (KAIST) has made a significant breakthrough in the field of artificial intelligence (AI) with the development of a cutting-edge semiconductor designed to process large language models (LLMs) with unprecedented efficiency.
The team, led by Professor Yoo Hoi-jun at the KAIST PIM Semiconductor Research Center, introduced the world’s first AI semiconductor, named the “Complementary-Transformer” AI chip. This innovative chip boasts the remarkable ability to process GPT-2 with ultra-low power consumption, utilizing just 400 milliwatts and achieving a high speed of 0.4 seconds, as reported by the Ministry of Science and ICT.
Illustration of AI semiconductor technology
Advancements in Power Efficiency and Size
The AI chip, measuring a mere 4.5-mm-square, was developed using Samsung Electronics’ 28 nanometer process, showcasing a remarkable 625 times reduction in power consumption compared to Nvidia’s A-100 GPU, a global AI chip leader. Nvidia’s A-100 GPU typically demands 250 watts of power to process LLMs, making KAIST’s semiconductor a game-changer in energy efficiency.
Moreover, the Complementary-Transformer chip is 41 times smaller in size than its Nvidia counterpart, enabling its integration into compact devices such as mobile phones, paving the way for enhanced AI capabilities in handheld technology.
Neuromorphic Computing Technology Integration
The successful development of this ultra-low power AI semiconductor was made possible through the integration of neuromorphic computing technology, specifically spiking neural networks (SNNs). While SNNs were previously known for their simplicity in tasks like image classification, the KAIST research team elevated their accuracy to match that of deep neural networks (DNNs), allowing for their application in processing LLMs.
The team’s innovative approach involves optimizing computational energy consumption by employing a unique neural network architecture that combines DNNs and SNNs. This fusion, coupled with effective compression techniques for large LLM parameters, ensures both energy efficiency and accuracy in AI processing.
In the words of the research team, their new AI chip represents a significant leap forward in AI technology, offering a groundbreaking solution that addresses the critical balance between computational power and energy efficiency in large language model processing.
As the field of artificial intelligence continues to evolve, innovations like KAIST’s ultra-low power LLM accelerator are poised to redefine the boundaries of AI capabilities, opening up new possibilities for efficient and powerful AI applications across various industries.
