A high throughput generative vector autoregression model for stochastic synapses.

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Computational Neuroscience

Background:

• Emerging electronic nanodevices mimic brain's synaptic plasticity for neuromorphic systems.
• Accurately modeling device response, hysteresis, and noise in synaptic arrays is crucial for large-scale simulations.

Purpose of the Study:

• To develop a high-throughput generative model for synaptic arrays.
• To accurately capture device parameters, temporal dynamics, and cross-correlations from real-world resistive memory cell data.

Main Methods:

• Utilized a vector autoregressive stochastic process to model device data.
• Mapped electrical measurement data from resistive memory cells onto the stochastic process.
• Developed parallelized implementations for CPUs and GPUs.

Main Results:

• The model accurately reproduces device parameters and their cross-correlation structure.
• Achieved high throughput exceeding one hundred million weight updates per second.
• Demonstrated scalability for array sizes exceeding one billion cells.

Conclusions:

• The developed generative model enables efficient and accurate large-scale simulations of neuromorphic systems.
• This approach overcomes key challenges in modeling complex device behaviors.
• The model's speed and accuracy pave the way for advanced neuromorphic hardware development.