CMOS + stochastic MTJ: Heterogeneous p-computers for energy-based machine learning

The nearing end of Moore's Law led to the rise of domain-specific hardware to extend capabilities of existing CMOS technology. Probabilistic computing with p-bits has emerged as a promising computing platform, naturally applicable to probabilistic models and algorithms [1,2]. Here, we present an experimental demonstration of a heterogeneous computer where stochastic Magnetic Tunnel Junctions (sMTJ) are connected to conventional Field Programmable Gate Arrays (FPGA). The sMTJs drive a large number of digital p-bits in the FPGA by providing asynchronous and truly random bits. We show how this platform can be used to train a class of physics-inspired, energy-based Machine Learning models. Compared to purely digital and pseudorandom number generators (PRNG), augmenting the FPGA with the truly random bits from the sMTJs increase the quality of randomness, resulting in superior performance in learning, while decreasing the area footprint and energy consumption needed from PRNGs. Our results highlight the promise of repurposing embedded magnetic memory technology to design energy-efficient and scalable CMOS + X architectures.

[1] W. A. Borders, et al., Nature 573, 390 (2019).

[2] Camsari, et al., Applied Physics Reviews, 6, 1 (2019).