From the Hailo Technology section and the compile step for models, we know that the Hailo-8 has three different types of resources: control, compute and memory. The purpose of memory is clear: store data. I assume the compute unit performs the calculations and the control unit handles the flow of data, schedules calculations, etc.
Memory resources store data—this includes both the model weights and any intermediate results that get generated while the neural network is processing.
Compute resources handle the actual number crunching needed for neural network operations.
Control resources (also called Layer Controller Units or LCUs) manage how data flows through the system, schedule operations, and coordinate everything happening on the device.