Thrust 1: Devices

1. Amongst the wide range of random phenomena exhibited by different devices, how can we match the randomness needed by different applications?
2. How can different computational approaches leverage the other unique characteristics of these devices?

To address these questions, we have been developing random number generators based on magnetic tunnel junctions and tunnel diodes. These two device technologies were chosen because of both devices produce large bistable signals that require minimum effort to condition to operate with conventional logical circuits, and they have manufacturing processes which readily integrate with silicon microfabrication techniques. Magnetic tunnel junctions (MTJs) measure the magnetization of a nanomagnet. First setting the nanomagnet into an unstable state, the MTJ produces well-distinguished signals that discriminate which of the two stable magnetizations the nanomagnet relaxes into. Tunnel diodes conduct current in two modes – either tunneling or thermionic emission – each having distinct voltage levels and current ranges. The detailed configuration of charge defects in the junction of the diode fluctuates with thermal energy, and determines which of the two modes the diode will operate in. 

In studying these devices, we are uncovering metrics that evaluate the quality of their statistical output, how these devices can be built up into efficient (size, speed, and power) circuits, and how to leverage their unique characteristics. For example, both the MTJ and the TD are tunable, meaning the probability of generating a heads or tails does not have to be fixed at 1:1. Weighted bitstreams can be used to produce samples from arbitrary distributions, and not just uniform distributions, which require additional calculations to transform to samples from other distributions.