Dynamic vision sensors (DVSes) allow catching pixel-wise motion in a scene with a us-level temporal resolution. As such, they now emerge as a stepstone toward ultra-low latency edge applications that include autonomous drone navigation, VR headsets and area surveillance. By forming a spatiotemporal graph of events, event-driven graph neural networks (GNNs) harness both geometrical and neighborhood information to process in real-time the extremely sparse information of DVSes, yielding a higher accuracy and a lower FLOPs/ev than CNN/SNN alternatives. Yet, conventional AI hardware accelerators fail to concurrently address the event-driven nature of the processing, the irregular memory accesses inherent to the sparse graph structure, and the regularity of the GNN's compute operations, thereby failing to harness their expected benefits.
In this talk, we investigate how us-level inference can thus be effectively leveraged through the careful HW-SW co-design of event-driven GNN accelerators. We start by showcasing how temporal graph causality can be exploited in HW to leverage a 16us/event inference on FPGA in small GNNs and with low-resolution images. Then, we point out critical scalability bottlenecks and propose HW-SW solutions to work around them. Lastly, we introduce ETHEREAL, a 28nm GNN chip accelerator tailored for state-of-the art asynchronous GNN workloads. ETHEREAL achieves a 0.15-10µs/inference at 250MHz and 150mW on a DAGr-S network, a performance notably enabled by its 3D-2D split memory hierarchy and its novel spatiotemporal caching strategy.
12/12/2025 11:00 - 12:00
ESAT Aula L
