Ultra-low-power probabilistic graphical computing based on memristors enabling communications signal processing
TL;DR
A probabilistic graphical computing paradigm based on memristors is reported, which not only unifies the framework of Bayesian inference-based baseband signal processing algorithms, but also achieves a substantial improvement in energy efficiency via memristor-based in-memory computing.
Abstract
Improving the energy efficiency of baseband processors has become a critical challenge in mobile communications, especially for energy-constrained edge devices. Applications in conventional baseband signal processors usually adopt independently optimized algorithmic architectures and primarily rely on digital systems, whose energy efficiency is fundamentally limited by frequent data movement and computationally intensive dynamic operations. To overcome the fundamental energy-scaling bottleneck in modern signal processing, here we report a probabilistic graphical computing paradigm based on memristors, which not only unifies the framework of Bayesian inference-based baseband signal processing algorithms, but also achieves a substantial improvement in energy efficiency via memristor-based in-memory computing. Results demonstrate that the proposed approach achieves energy consumption below 10 pJ per bit and up to an 11.4-fold improvement in energy efficiency over conventional implementations, while maintaining error rates below 1×10−4 in the representative application. These results provide a flexible, noise-resilient route to ultra-low-power signal processing in edge devices. To address the energy bottleneck in modern mobile communications, Jiang et al. introduce a memristor-based framework unifying key wireless signal-processing tasks. It achieves sub-10 pJ/bit consumption and an 11.4-fold efficiency gain while maintaining low error rates for edge communications.
