国产RRAM：闪耀ISSCC

Core Insights - The article highlights the significant achievements of Hefei Reliance Memory (合肥睿科微电子) at the ISSCC 2026, showcasing China's advancements in RRAM technology and its potential in AI applications [2][19] - The company has developed a unique ReRAM-based chip that addresses key challenges in AI inference, particularly in large language models (LLMs) and edge AI applications, demonstrating a successful integration of mature process technology with innovative architecture [3][4][17] Group 1: RRAM Technology and Innovations - Hefei Reliance Memory has leveraged its proprietary ReRAM technology to create a chip that excels in performance, energy efficiency, and cost-effectiveness, suitable for AI inference across various scenarios [3][4] - The ReRAM technology offers advantages such as non-volatility, high-speed read/write capabilities, low power consumption, and high-density integration, making it a promising solution for AI, IoT, and edge computing [4][18] - The company’s approach combines mature 55nm CMOS processes with innovative architecture, allowing it to challenge advanced process technologies effectively [4][17] Group 2: Breakthroughs in AI Inference - The company introduced a 55nm LLM accelerator that utilizes a 3D stacked ReRAM-on-Logic architecture, significantly improving performance and reducing costs for AI model inference [9][10] - Key innovations in the accelerator include a locally rotating unit that enhances inference speed by 3.82-3.93 times while saving 92.7% chip area, and a stacked architecture that eliminates external memory access delays [10][11] - The accelerator achieves a peak performance of 2.33 TOPS with a power consumption of only 49.54mW per ReRAM chip, demonstrating its capability to meet the demands of LLM inference [11][12] Group 3: Edge AI Applications - A fully analog intelligent vision SoC developed in collaboration with local research teams has achieved significant efficiency in edge AI applications, marking a breakthrough in end-to-end processing without the need for A/D conversion [14][16] - This SoC integrates various components, including a PWM image sensor and ReRAM memory, to deliver high-speed, low-power performance suitable for edge devices [15][16] - The chip's energy efficiency reaches 345.54 TOPS/W, with substantial improvements over previous solutions, making it ideal for applications in smart wearables and autonomous driving [16][18] Group 4: Industry Implications - The advancements made by Hefei Reliance Memory not only signify a technological leap for China in the semiconductor industry but also provide a viable path for domestic AI inference solutions, reducing reliance on foreign technologies [17][18] - The combination of ReRAM technology with mature processes aligns with the industry's trend towards cost reduction and efficiency, supporting the broader goal of achieving self-sufficiency in semiconductor manufacturing [18][19] - The success at ISSCC 2026 serves as a model for other domestic companies to follow, emphasizing the importance of innovation and collaboration in advancing the semiconductor landscape [17][19]