Memristive In-Memory Object Detection with 128 Mb C-Doped Ge₂Sb₂Te₅ PCM Chip.

Area of Science:

• Computer Vision
• Materials Science
• Electrical Engineering

Background:

• Traditional edge computing struggles with the computational demands of modern object detection models.
• State-of-the-art object detection requires substantial computing power and memory, limiting edge device capabilities.

Purpose of the Study:

• To develop an efficient in-memory computing system for object detection using phase change memory.
• To address the limitations of traditional computing platforms for AI at the edge.

Main Methods:

• Fabrication of a 128 Mb phase change memory chip.
• Implementation of in-memory vector-matrix multiplication and max computation.
• Adoption of a mixed-precision weight mapping strategy to reduce energy consumption and reliance on high-precision analog-to-digital converters (ADC).

Main Results:

• Achieved a 99.99999% memory yield in a 40 nm node.
• Reduced ADC energy consumption by up to 22.3× with equivalent network performance.
• Demonstrated 4,180× higher energy efficiency and 228× greater computational throughput compared to GPU implementations.

Conclusions:

• Memristive in-memory computing offers a highly efficient solution for object detection at the edge.
• The developed system significantly surpasses GPU performance in terms of energy efficiency and speed.
• Mixed-precision strategies are crucial for optimizing compute-in-memory operations.