Abstract
Neuromorphic systems inspired their design from the mammalian biological brain by using spiking neurons as processing elements and synapses as memory. Unlike the conventional architecture, neuromorphic systems have more parallelism, higher throughput, and lower power consumption thanks to their processing elements’ asynchronous and event-driven behavior. To map the 3D structure of the brain, using the 3D integrated circuits (3D-ICs) and 3D network-on-chip (3D-NoC) interconnect is more suitable for scalable design, lower power consumption, and shorter wires. However, 3D-NoC suffers from thermal issues due to the high power density in the small area stacked dies and the heterogeneous heat transfer between the layers. This paper presents a thermal-aware mapping method for 3D-NoC-based neuromorphic systems that moderates the temperature between the layers. Our approach is divided into two stages. First, we cluster the neurons, and second, we place the clusters in their respective tiles. The results show that the proposed mapping technique moderates the temperature of the tiles across each layer where the temperature variance between the tiles of a layer is up to only $1.5^{\circ}\mathrm{C}$ in a 3D-NoC. In contrast, for random mapping, it reaches $9.7^{\circ}\mathrm{C}$.

This publication has 30 references indexed in Scilit: