A compute-in-memory chip based on resistive random-access memory.

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Edge devices require energy-efficient hardware for complex AI functionalities.
• Compute-in-memory (CIM) using resistive random-access memory (RRAM) offers a solution by integrating memory and computation.
• Existing RRAM-CIM chips face challenges in balancing energy efficiency, model versatility, and accuracy.

Purpose of the Study:

• To develop a RRAM-based CIM chip, NeuRRAM, that overcomes the trade-offs between efficiency, versatility, and accuracy.
• To demonstrate simultaneous improvements across multiple design hierarchies: algorithms, architecture, circuits, and devices.
• To enable advanced AI functionalities directly on edge devices with unprecedented energy efficiency.

Main Methods:

• Co-optimization across algorithms, architecture, circuits, and devices for RRAM-CIM design.
• Development of a novel RRAM-based CIM chip named NeuRRAM.
• Integration of dense, analog, non-volatile RRAM devices for in-memory computation.

Main Results:

• NeuRRAM achieves two-times greater energy efficiency compared to prior RRAM-CIM chips.
• The chip demonstrates versatility by reconfiguring CIM cores for diverse AI model architectures.
• Inference accuracy is comparable to software models with four-bit weight quantization across various AI tasks, including image classification and speech recognition.

Conclusions:

• NeuRRAM represents a significant advancement in RRAM-based CIM technology.
• The co-optimization approach successfully addresses the efficiency-versatility-accuracy trade-offs.
• This technology paves the way for highly efficient and accurate AI processing on edge devices.