温控新技术显著提升AI核心硬件性能

Core Insights - Scientists from Sungkyunkwan University have developed a new technology that utilizes heat to precisely control the internal structure of semiconductors, significantly enhancing the performance of next-generation AI core hardware, enabling complex AI computations to be processed faster with lower power consumption [1][2] Group 1: Technology Development - The current architecture of computers separates processors and memory, leading to inefficiencies as data must be transferred back and forth, akin to students moving between a desk and a bookshelf [1] - The proposed solution, "in-memory computing," allows computations to be performed directly within the memory, focusing on the use of ferroelectric transistors [1] - The key challenge in manufacturing these transistors lies in the material hafnium oxide, which requires specific atomic arrangements to maintain storage functionality [1] Group 2: Innovative Approach - Previous methods to address the challenges involved doping with other chemical elements, which are complex and not suitable for mass production [1] - The research team employed the principle of thermal expansion, where different materials expand and contract at varying rates when heated, to overcome these challenges [2] Group 3: Performance and Applications - The team designed electrodes around the semiconductor material that apply compressive stress to the hafnium oxide during cooling, effectively "shaping" the atoms into an optimal crystalline structure for storage operations [2] - Tests indicate that the newly designed semiconductor devices remain stable after over one trillion operations and achieve an accuracy rate of 97.2% in AI image recognition tasks [2] - The results suggest that high-performance AI semiconductor devices can be achieved without complex chemical processes, relying solely on temperature control, which could lead to smarter and more efficient AI technologies in applications like autonomous vehicles and smartphones if commercialized [2]