The team from the University of Hong Kong, Hong Kong University of Science and Technology, and Xi’an University of Electronic Science and Technology has published results in a Nature sub-journal, utilizing memristors to create an adaptive ADC, reducing the power consumption of AI chips by 57.2% and the area by 30.7%. [Introduction] The team from HKU, HKUST, and Xidian University has made a breakthrough in addressing the core challenges of AI chips, tackling the analog-to-digital converter (ADC), which accounts for 87% of energy consumption in the compute-in-memory (CIM) architecture, a significant “energy black hole.” By leveraging the programmable characteristics of memristors, they have developed a “smart ruler” that adapts to data distribution, resulting in a 57.2% reduction in power consumption and a 30.7% decrease in area for AI chips, paving the way for the next generation of efficient AI hardware systems. In today’s context of exponentially increasing demand for AI computing power, the “compute-in-memory” (CIM) architecture is seen as the future of AI inference chips. It utilizes fundamental physical laws to achieve computation, offering significant energy efficiency advantages over GPUs. However, a critical “energy black hole”—the analog-to-digital converter (ADC)—has severely hindered its development. In advanced CIM chips, the ADC can consume up to 87.8% of energy and 75.2% of area, greatly suppressing the potential that CIM AI chips should have. Recently, a joint research team led by Liu Zhengwu, Zhang Wei, Li Can, and Huang Yi from the University of Hong Kong, Hong Kong University of Science and Technology, and Xi’an University of Electronic Science and Technology has directly tackled this issue. The first author of the paper, Hong Haiqiao, has proposed for the first time internationally a hardware-native adaptive ADC architecture based on memristors. Paper link: https://www.nature.com/articles/s41467-025-65233-w Code link: https://github.com/MIKEHHQ/ReADC This design innovatively utilizes the programmable characteristics of memristors, making the ADC “ruler” smart and efficient, significantly reducing the energy overhead of the ADC module in CIM chips by 57.2% and decreasing the area by 30.7%, thus paving the way for the next generation of efficient AI hardware systems.
……
Link to detailed content:
[Nature Sub-Journal: Next-Generation AI Hardware System Proposed by HKU and Others, Energy Consumption Reduced by 57.2% – CSDN App]