The third option between a conventional von Neumann machine and true in-memory compute. Keep the compute general and digital, but move the memory physically closer to it, so the data has less far to travel. The compute and the memory stay separate functions; only the distance changes.
How it differs from its neighbours
The shared goal is to attack the Von Neumann Bottleneck and the The Memory Wall by cutting data movement; near-memory does it with packaging rather than with a new device.
Why it matters here
Near-memory is the lowest-risk, most-productised answer (it uses standard cells and standard memory), which is exactly why it commoditises: 3D-stacked SRAM-on-logic is becoming a foundry service. It is a key category in the Hbm Free Inference Architectures taxonomy and the foil against which in-memory and charge-domain approaches have to justify their device risk.