ELSA: A Near-SRAM Architecture for Elastic SNN Inference
A new paper on arXiv (2605.20802) introduces ELSA, a near-SRAM dataflow architecture designed to enable true elastic inference in spiking neural networks (SNNs). SNNs offer event-driven, addition-only computation for improved efficiency, and their elastic inference property allows outputs to emerge progressively, enabling early responses to salient inputs. However, existing SNN accelerators—whether layer-by-layer or time-step-by-time-step—cannot exploit this property due to coarse-grained pipelines that delay output forwarding. ELSA overcomes this with a fine-grained spine/token-wise pipeline and hardware optimizations, allowing immediate forwarding of results. The architecture is tailored for near-SRAM integration, aiming to realize the full benefits of elastic inference in neuromorphic computing.
Key facts
- ELSA is a near-SRAM dataflow architecture for SNN inference.
- It enables true elastic inference through fine-grained spine/token-wise pipeline.
- Existing SNN accelerators use layer-by-layer or time-step-by-time-step designs.
- These designs delay output forwarding and forfeit elastic inference benefits.
- ELSA allows immediate forwarding of results for early responses.
- The paper is available on arXiv with ID 2605.20802.
- SNNs exploit event-driven and addition-only computation.
- Elastic inference allows outputs to emerge progressively.
Entities
Institutions
- arXiv