Architecture
The system integrates an AXI4-Lite control path, arbiter, controller, 1024 × 64-bit instruction memory, and a 4 × 4 processing-element array.
Each processing element contains a 32-bit ALU, load-store unit, instruction register, dual data outputs, and 1024 × 32-bit local memory.
Data movement
A custom 64-bit instruction format and programmable row connections configure both PE operations and data routing without redesigning RTL for each kernel.