Abstract:
This article presents a modular 3D stacked manycore processor that integrates a general-purpose RISC-V base die with a domain-specific deep neural network (DNN) accelerator die using a 9 $\mu $ m hybrid-bonded interconnect (HBI). A custom $14 \times 4~\times 2$ 3D network-on-chip (3D NoC) provides a flat, distributed shared-memory address space across 112 tiles, enabling …