Presentation
Late Breaking Results: A Neuromorphic Accelerator for Efficient Multi-Head Attention Processing in Spiking Vision Transformers
DescriptionSpiking Vision Transformers (SViTs) are developed as an energy-efficient alternative to conventional ViTs. To maximize efficiency gains of SViT processing, we propose MorphAtt, a novel digital SViT accelerator that expedites the inference through streamlined processing. Specifically, it processes multi-head self-attention (MHSA) operations using spiking Query-Key-Value, spiking self-attention, and reparameterization modules, as well as inter-module buffers to mitigate traffic congestion in on-chip memory accesses. Experimental results show that, MorphAtt achieves 792-1605 GOPS of throughput, while incurring ∼39-55 mW of power consumption and 1.5 mm2 of area, which lead to 20.3-29.1 TOPS/W of energy efficiency. These results demonstrate that, our MorphAtt offers better performance and efficiency trade-offs than state-of-the-art, thus enabling highly energy-efficient vision-based systems at the edge.
Event Type
Work in Progress
TimeMonday, July 275:27pm - 5:27pm PDT
LocationExhibit Hall
Similar Presentations