A full-stack memristor-based computation-in-memory system with software-hardware co-development

Area of Science:

• Computer Engineering
• Materials Science
• Artificial Intelligence

Background:

• Memristor-based computation-in-memory (CIM) systems face limitations in practicality due to hardware design constraints and manual parameter tuning.
• Optimizing CIM systems is crucial for efficient and robust artificial intelligence hardware.

Purpose of the Study:

• To develop a software-hardware co-development approach for enhancing the flexibility and efficiency of memristor-based CIM systems.
• To improve the robustness of neural network models against hardware non-idealities and analogue computing noise.

Main Methods:

• Implemented a flexible hardware component supporting diverse dataflow and mapping strategies.
• Developed software for automatic model placement and efficient optimization techniques.
• Integrated software and hardware for a full-stack co-development solution.

Main Results:

• Demonstrated the system on six neural network models across four tasks.
• Achieved a 4.76% accuracy improvement for ResNet-32 during training.
• Observed 3.32% to 9.45% accuracy improvements across models during on-chip inference.

Conclusions:

• The proposed software-hardware co-development approach significantly improves the performance and practicality of memristor-based CIM systems.
• Automatic optimization methods enhance model robustness and suppress noise, leading to higher accuracy in AI applications.