Energy-efficient stochastic computing with superparamagnetic tunnel junctions

We design an efficient stochastic bitstream generator based on superparamagnetic tunnel junctions, which can produce low energy, truly random bits, in turn drastically reducing cross-correlation. This bitstream generator allows us to address an outstanding issue in stochastic computing: that it has been limited by the inaccuracies introduced by correlations between the pseudorandom bitstreams used in the calculations. This bitstream generator gives us the freedom of not having to design around correlations and allows us to propose a low-energy approach to stochastic computing. To demonstrate the effectiveness of this approach, we incorporate it into an efficient CMOS neural network design. Our simulations of this network reach error rates comparable to recent work in stochastic-computing-based neural networks at nearly an order of magnitude lower energy expenditure.