Reproduce “ETHEREAL: A 25.6-$μ$s/inf. Low-latency Event-driven Graph-neural-network Processor for High-resolution Vision at the Edge” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Adrian Kneip, Martin Lefebvre, Daniel Gehrig, Victoria Catalán Pastor, Davide Scaramuzza, Marian Verhelst, Charlotte Frenkel
ETHEREAL, a specialized processor, accelerates Event-driven Graph Neural Networks (EV-GNNs) for high-resolution edge vision, demonstrating a 25.6 μs/event inference latency and 1.6 μJ/event energy consumption on a 640x480 pixel workload. Its design addresses the challenges of mixed dense-regular compute and sparse-irregular memory access patterns inherent in EV-GNNs.
Key insights
- Reformulating spline convolution into a three-phase process and enabling neighbor-level parallelism improves execution time by up to 3.8x for complex operations with flexible precision.
- A tiered memory hierarchy, featuring a 3D spatiotemporal cache and a banked 2D spatial scratchpad, effectively manages sparse and dense graph data, reducing external memory accesses by 30-60% and 2D layer latency by 4.3-7.3x respectively.
- Pipelining graph-building for subsequent events with current event processing masks graph-building latency, leading to a 1.8x reduction in overall end-to-end inference latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.