Energy-Efficient Float16 Sampling with Stochastic Magnetic Tunnel Junctions
Configuring an array of current-biased stochastic magnetic tunnel junction (s-MTJ) devices as tunable per-bit Bernoulli generators mapped onto the sign/exponent/mantissa positions of the Float16 format, so repeated hardware draws converge to a uniform floating-point distribution; arbitrary 1D distributions are then sampled via a mixture-of-uniforms representation with convolution and prior-likelihood operations. The device-native, symbolic-computation-free approach yields roughly 5649x lower energy than PCG and 9721x versus Mersenne-Twister while providing genuine (thermal-noise) randomness for probabilistic ML and MCMC.
Energy-Efficient Sampling Using Stochastic Magnetic Tunnel Junctions
(Pseudo)random sampling, a costly yet widely used method in (probabilistic) machine learning and Markov Chain Monte Carlo algorithms, remains unfeasible on a truly large scale due to unmet computatio…