Energy Efficient p-Circuits for Generative Neural Networks

Abstract: A p-circuit comprising a network of $N$ p-bits can generate samples from a probability distribution with $2^N$ possibilities just like a network of q-bits, though there are interesting fundamental differences. A p-bit is essentially a binary stochastic neuron used in Boltzmann machines which involve the repeated application of a specific elementary operation, whereby each p-bit looks at the state of 'n' other p-bits, computes the probability for being $0$ or $1$, and chooses its next state accordingly. We argue that the energy cost, $\epsilon(n)$, of this elementary operation is an important metric and we use our ASIC to extract it for our implementation, to serve as model and benchmark for other implementations. We further argue that the same elementary operation can be used in generative AI replacing the standard continuous or multi-bit digital neurons with p-bits. It might appear that this would introduce significant errors, but we show that by averaging a small number (say $1 \ to \ 100$) samples, the image quality can be restored.