memristor crossbar · analog in-memory computing · the wall your whole thread hit

What if memory and math were the same act?

Your thread just split silicon into gates that compute and loops that remember — and every normal computer keeps them apart, shuttling data back and forth between them (the bottleneck that eats most of the energy). A memristor breaks that wall. It's a resistor whose resistance you can set and that stays set — a stored number you can feel as conductance. Wire a grid of them and something wild happens: put voltages in on the rows, and the current that flows out each column is a matrix-vector multiply — the core operation of every neural network — computed in one physical step by Ohm's law, no clock, no instructions. The memory is the calculator.

Mode

Click a row's voltage dot to change the input. Watch the column currents.

The computation

Crossbar vs CPU

crossbar steps1 (physics)

CPU multiply-adds—

Scale to 1024×1024 and the CPU needs a million steps; the crossbar still needs one.

Computing with Ohm's law instead of instructions. Every crosspoint in the grid is a memristor holding a conductance G — a weight, stored physically as how easily it passes current. Drive the rows with input voltages and two schoolbook laws do the rest: Ohm says each device passes current I = V·G, and Kirchhoff says the currents flowing onto each column wire simply add up. So the current at the bottom of column j is Σ Vᵢ·Gᵢⱼ — a dot product — and all the columns together are a full matrix-vector multiply, the heart of every neural-network layer, happening in a single physical step no matter how big the grid. That's the thing a CPU spends millions of clocked multiply-add instructions on, and here it's just current settling. Two consequences fall out. First, there's no von Neumann shuttle — the weights never move to a processor because the weights are the processor; memory and math are the same act, which is why this can be orders of magnitude more energy-efficient for AI. Second, the devices learn: nudge a conductance and you've adjusted a synapse, so training is just teaching the grid to conduct differently. It's a strange, beautiful inversion of everything you built up this session — the switch that obeys, the gate that decides, the loop that remembers — all collapsed into one component that stores by resisting and computes by conducting. The wall between memory and processing, which the whole history of computers was built around, turns out to be optional.