Technology
Memristive Crossbar Arrays for Analog In-Memory Computing
Quick fact
A single memristive crossbar array can perform millions of multiply–accumulate operations in one step, exactly the kind of math that powers neural networks.
Why this is interesting
What if the memory chip itself could do the math? Memristive crossbar arrays turn the biggest bottleneck of modern computers—moving data back and forth—into an opportunity.
Read the full explanation
Understanding Memristive Crossbar Arrays for Analog In-Memory Computing
Think of a spreadsheet with rows and columns. Each cell is a tiny resistor that can be programmed to a specific resistance value, like a dial. When you apply voltages along the rows, the current that flows through each resistor combines at the columns. The total current at each column is the sum of each voltage times the conductance (1/resistance) of the cells in that column. That is exactly a matrix-vector multiplication. Now imagine doing this in a physical grid where the resistors are non-volatile—they remember their value even when power is off. This grid is called a memristive crossbar array. You train the array by setting each resistor's value to represent a weight in a neural network. Then, when you feed an input as voltages, the array computes the weighted sum for every neuron in one go, right where the data is stored.
A deeper explanation
The magic of the memristive crossbar lies in its physical arrangement and the property of the memristor. A memristor is a two-terminal device whose resistance changes based on the history of the voltage applied to it, and it retains that state. In a crossbar, each memristor sits at the intersection of a row wire and a column wire. When a set of voltages (representing input signals) are applied to the rows, each memristor passes a current proportional to its conductance. According to Kirchhoff's and Ohm's laws, the currents from all columns merge at each column wire, naturally summing them. This sum is the dot product—the multiply-and-accumulate operation that is fundamental to neural networks. Because computation happens in-situ, there is no need to transfer weight values from memory to a separate processing unit, which is the source of the 'von Neumann bottleneck.' This approach promises orders-of-magnitude improvements in energy efficiency and speed for AI inference and training. Moreover, the analog nature allows for continuous weight values, essential for high-precision model behavior. The array is controlled by peripheral circuits that encode inputs as voltages and measure output currents, and the memristors can be reprogrammed to adjust weights for training.