185 femtojoules per cell per operation at 100 MHz. That is the energy cost of one TXL-RBF neuron classifying a handwritten digit, and it makes most edge ML accelerators look like space heaters.
The authors of the paper behind arXiv:2606.14739 built a Radial Basis Function classifier using Metal-Oxide Resistive RAM (RRAM) arrays wired as Analog Content Addressable Memory (ACAM). Their trick: a custom Template piXeL (TXL) cell that acts as a configurable receptive field neuron, directly computing the distance between an input and a stored prototype in the analog domain.
185fJ Per Cell Per Operation at 100MHz
Each TXL cell embeds a Radial Basis activation function. Apply a voltage representing an input pixel, and the cell outputs a current proportional to the Gaussian distance from its programmed center. Stack these cells into dense arrays and you get a full similarity search engine that runs in a single clock cycle, no multiply-accumulate necessary.
The entire classifier hit 89.1% accuracy on the MNIST test set. Not ground-breaking for MNIST in 2025, but consider the power: 185fJ per cell per operation. That is roughly 70% less energy than a typical SRAM-based CAM solution at the same technology node, because RRAM cells are non-volatile and the analog computation happens at the memory itself.
How TXL Cells Turn RRAM Into a Receptive Field Engine
Instead of streaming data to a distant processor, the TXL array computes distances locally. Each cell stores its receptive field center as an RRAM resistance value. Input voltages are applied across the cell, and the resulting current encodes the similarity. No ADC, no digital multipliers, no Von Neumann bottleneck.
The ACAM architecture lets you program prototypes on the fly. Change the RRAM resistance distribution, and the receptive field shifts. That enables what the paper calls on-the-fly learning: retune the classifier to handle domain drift without pausing inference.
Online Learning Without the Cloud
Safety-critical edge applications like autonomous navigation cannot afford to phone home for model updates. The TXL-RBF design lets the hardware adapt its decision boundaries as the environment changes, using local feedback to adjust RRAM states. The authors claim this works for incremental learning, though they only demonstrate static classification on MNIST so far.
What this really enables is a class of extreme-edge devices that can learn and classify for years on a coin cell battery. 185fJ per cell per op means a 100 mAh battery could support over 1.9 trillion cell operations. For a 784-pixel MNIST input, that is about 2.4 billion full classifications per battery charge. That number changes the math on where you can deploy machine learning.
Source: An RRAM-based Hardware Implementation of a Radial Basis Function Neuron for Edge Classifiers
Domain: arxiv.org
Comments load interactively on the live page.