A systolic array is a grid of processing units that pass data to their neighbors in step, like a pulse. Each unit performs a small calculation, usually a multiply-accumulate, and hands its result on, so data flows through the grid without being fetched from memory again at every step.
This design suits matrix multiplication well, which is why it appears in many AI accelerators and NPUs. Reusing data inside the array reduces memory traffic and energy use compared with fetching the same values repeatedly.
A systolic array is not the only way to build such hardware, and having one does not mean a chip is better. It works best on large, regular calculations and can be less efficient on small or irregular ones. Real speed still depends on how the software maps work onto the array and on memory bandwidth.
When you read about one, treat it as an explanation of how a chip achieves its efficiency, not as a performance claim in itself. Compare chips on measured results for the workloads you care about.