Dynamic vision sensors generate brightness-change events with microsecond time resolution, which is why robotics and autonomy teams adopted them. The problem: until now there was no hardware capable of processing those event streams fast enough to actually use that temporal resolution. Conventional vision accelerators are designed for dense frame buffers; event camera outputs are sparse, irregular spatiotemporal graphs. ETHEREAL is the first chip designed specifically for event-driven GNN inference. At 25.6 microseconds per inference, it processes events faster than the inter-event interval at production sensor resolutions.
The architecture makes a clean design choice: no frame reconstruction step, no conversion to dense tensors. A neighbor-parallel spline-convolution engine handles the sparse GNN compute directly, paired with a split 2D/3D memory hierarchy that caches spatiotemporal event neighborhoods rather than pixel grids. This is end-to-end event-native compute, which is why the latency drops below what any frame-conversion pipeline can achieve.
Teams building robotic manipulation, visual odometry, or high-speed industrial inspection who moved to event cameras but kept conventional vision inference hardware are leaving the sensor's core advantage unused. The DVS sees at microsecond resolution; the processing chain could not keep up, so the effective latency was still dominated by the inference step. ETHEREAL removes that asymmetry. The next constraint shifts to the data interface between sensor and processor, not the inference latency itself. If event cameras are on your roadmap for 2027 systems, the inference silicon problem is now solved.