Neuromorphic Processing Element Array

The AI accelerator is built around a 2-D mesh of identical processing elements (PEs), each pairing an L1 Spad SRAM with a tunable arithmetic pipeline whose multiplier bit-width, adder bit-width, and vector length can be set per workload. PEs exchange data over short, pipelined links across the mesh, while a centrally positioned, multi-ported L2-class global buffer streams weights and feature maps to the array and gathers partial results. This tiered memory hierarchy—private L1 Spad for high-reuse data and a shared L2 for larger tensors—minimizes off-chip traffic and lets the accelerator scale smoothly from energy-efficient INT8 inference to higher-precision operations simply by retargeting the per-PE arithmetic parameters without altering the mesh fabric or software stack.We also optimize dataflow and throughput by clustering multiple processing element during model compilation time.^{[More detailed models]}