Fusion of memristor and digital compute-in-memory processing for energy-efficient edge computing

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• AI edge devices require high-capacity, nonvolatile compute-in-memory (CIM) for energy efficiency and rapid response.
• Existing CIM solutions face trade-offs: memristor-based CIMs have limited endurance and accuracy, while SRAM-based CIMs are volatile and require large areas.

Purpose of the Study:

• To develop an AI edge processor that overcomes the limitations of existing CIM technologies.
• To leverage the strengths of both memristor and SRAM technologies for improved AI edge computing.

Main Methods:

• A novel memristor-SRAM CIM-fusion scheme was designed and implemented.
• The fusion processor integrates the accuracy of SRAM CIM with the energy efficiency and density of memristor CIM.
• The system supports adaptive local training for personalized AI capabilities.

Main Results:

• The fusion processor achieved high CIM capacity and a short wakeup-to-response latency of 392 microseconds.
• Achieved peak energy efficiency of 77.64 teraoperations per second per watt.
• Demonstrated robust accuracy with less than 0.5% accuracy loss and confirmed manufacturability of memristor technology for AI edge processors.

Conclusions:

• The memristor-SRAM CIM-fusion scheme offers a superior solution for AI edge devices, balancing accuracy, efficiency, and capacity.
• This technology enables on-device adaptive training, enhancing personalization and user experience.
• Memristor technology is now manufacturable and ready for integration into commercial AI edge processors.