Hardware

We develop circuits that generate and control randomness for probabilistic computing. Our building blocks include p-bits, whose binary states fluctuate with electrically controlled probabilities, as well as devices that produce tunable Gaussian random signals. We study how these elements behave when connected into programmable systems.

CMOS+X

Our CMOS+X approach combines CMOS circuits with stochastic devices, including magnetic tunnel junctions. We also study single-photon avalanche diodes (SPADs) as entropy sources and compare these approaches with fully CMOS implementations. Collaborative work includes a custom CMOS chip supplied by a separate magnetic entropy source.

Stochastic magnetic tunnel junction circuits supply asynchronous clocks to an FPGA in a CMOS+X probabilistic computer.
Stochastic magnetic tunnel junction circuits supply asynchronous clocks to an FPGA in a CMOS+X probabilistic computer.

CMOS+X hardware, cropped from Singh et al., Figure 1(e,f), CC BY 4.0. Full-resolution figure.

Learning with physical circuits

Fabricated circuits differ from their ideal models because their devices and interconnections vary. We develop CMOS probabilistic chips whose measured outputs guide the learning of weights and biases, helping compensate for circuit nonidealities. We also study optimization methods that account for the measured response of SPAD-based p-bits.

Digital probability control

DAC-free p-bits use digital control to tune output probabilities without a digital-to-analog converter at each p-bit.

Comparison of a conventional p-bit with a digital-to-analog converter and a DAC-free p-bit with digital probability control.
Comparison of a conventional p-bit with a digital-to-analog converter and a DAC-free p-bit with digital probability control.

Digital probability control without a DAC per p-bit. Selcuk et al., DAC-Free p-bits. Image: Shunsuke Fukami, Kerem Camsari et al., via Tohoku University. Full-resolution figure.

Measuring the cost of computation

Coupling stochastic elements, storing weights, and controlling probabilities require circuitry that adds area and consumes energy. We use measured prototypes and circuit models to examine these costs alongside sampling speed and statistical accuracy. These measurements guide integrated designs and identify which device improvements can benefit the complete system.

Foundational Publications