Thermal Integrity of ReRAM-based Near-Memory Computing in 3D Integrated DNN Accelerators

In this paper, the thermal integrity of near-memory computing in 3D integrated deep neural network (DNN) accelerators is investigated. Both conventional memory technologies (such as SRAM and DRAM) and emerging resistive memory (ReRAM) technology are considered. Through silicon via (TSV) based 3D integration and monolithic inter-tier via (MIV) based 3D technologies are leveraged for near-memory computing. The results demonstrate that monolithic 3D integrated DNN accelerator is thermally more feasible than TSV-based 3D accelerator due to reduced thermal resistance to heat sink. Furthermore, for near-memory computing with three-tier 3D systems, ReRAM based accelerator produces the lowest temperature whereas embedded DRAM (eDRAM) significantly increases the peak temperature.