CMOS plus stochastic nanomagnets enabling heterogeneous computers for probabilistic inference and learning.

Area of Science:

• Computer Engineering
• Materials Science
• Artificial Intelligence

Background:

• Moore's Law is being extended by integrating emerging nanotechnologies with CMOS transistors.
• Probabilistic machine learning, optimization, and quantum simulation often rely on Monte Carlo algorithms.
• Stochastic magnetic tunnel junctions (sMTJs) offer a promising avenue for novel computing paradigms.

Purpose of the Study:

• To develop an energy-efficient prototype combining CMOS technology with sMTJ-based probabilistic bits (p-bits).
• To demonstrate the capability of this hybrid system for probabilistic inference and learning.
• To evaluate the performance and energy efficiency of sMTJ p-bits compared to traditional CMOS transistors.

Main Methods:

• Integration of sMTJ-based p-bits with Field Programmable Gate Arrays (FPGAs).
• Utilizing asynchronously driven CMOS circuits controlled by sMTJs.
• Leveraging the algorithmic update-order-invariance property of Gibbs sampling for probabilistic inference.
• Augmenting low-quality random number generators (RNGs) with sMTJ stochasticity.

Main Results:

• A functional CMOS + sMTJ prototype was successfully created.
• The system demonstrated effective probabilistic inference and learning capabilities.
• sMTJ p-bits were shown to replace up to 10,000 CMOS transistors with two orders of magnitude less energy dissipation.
• The stochasticity of sMTJs can enhance the quality of random number generation.

Conclusions:

• The developed CMOS + sMTJ prototype offers a significant advancement in energy-efficient probabilistic computing.
• This approach can enhance the performance of deep Boltzmann machines and other energy-based learning algorithms.
• The integration of sMTJs with FPGAs paves the way for high-throughput and energy-efficient probabilistic machine learning and simulation.