Figure 10.

Tiny YOLO CNN architecture used for the object detection task (Gen4 dataset), along with the details on how the mapping has been done on SENECA cores and information about memory and energy consumption. The memory usage per core includes the total occupancy of the data memory (instruction memory consumption is about 20KB for all cores). Additionally, memory used for neuron states (16b per neuron) and parameters (weights and biases, 8b) are reported separately. The energy consumption is averaged over the entire dataset to infer a 50ms event stream and is presented in four columns: “All” shows the total energy consumption of the core, while “RV” is RISC-V and its instruction memory, “DM” is Data Memory, and “NP” is all the Neural Processing accelerators: NPEs, loop controller and event generator.