Researchers including UZH's Robotics and Perception Group demonstrated ETHEREAL, a processor designed for graph neural networks that analyse event-camera signals. Measurements on the DAGr-GNN workload and the 640-by-480-pixel DSEC dataset showed 25.6 microseconds of latency and 1.6 microjoules of energy per event-wise inference.
The chip combines parallel processing of neighbouring events with a memory design that caches relevant spatial and temporal information. It addresses the irregular memory access that makes these networks difficult to accelerate. The result is a measured research-chip demonstration, not a commercially available vision system; the August manuscript was submitted for possible journal publication.
