Energy Scaling Advantages of Resistive Memory Crossbar Based Computation and Its Application to Sparse Coding

Area of Science:

• Computer Science
• Artificial Intelligence
• Materials Science

Background:

• The exponential growth of data necessitates advanced analytics solutions.
• Biological systems offer insights into low-power, efficient data processing.
• Analog resistive memory crossbars present a promising hardware platform for neuromorphic computing.

Purpose of the Study:

• To explore the energy efficiency of analog resistive memory crossbars for data analytics.
• To evaluate the application of crossbar kernels in neuromorphic algorithms.
• To demonstrate the potential of these systems for tasks like sparse coding.

Main Methods:

• Utilizing analog resistive memory crossbars for vector-matrix multiplication (read) and rank-1 update (write) operations.
• Analyzing the energy efficiency of these operations compared to conventional digital architectures.
• Applying the crossbar kernels to a neural sparse coding algorithm.

Main Results:

• Analog crossbars achieve O(N) energy efficiency for vector-matrix multiplication and rank-1 updates.
• Read operations can achieve O(1) energy efficiency under noise-limited conditions.
• Application to sparse coding yields an O(N) energy reduction for the algorithm.

Conclusions:

• Analog resistive memory crossbars offer a highly energy-efficient hardware solution for data-intensive applications.
• The demonstrated kernels are fundamental to various neuromorphic algorithms, including image, text, and speech recognition.
• This approach significantly advances the feasibility of low-power, high-performance AI and machine learning systems.