Hardware accelerator for vision transformers using dynamic token pruning, ReLU replacement, FFN pruning, and row-wise dataflow to reach 2.31 TOPS/W in 28nm CMOS with under 2% accuracy loss.
ELSA: hardware-software co-design for efficient, lightweight self- attention mechanism in neural networks
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AR 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Low Power Vision Transformer Accelerator with Hardware-Aware Pruning and Optimized Dataflow
Hardware accelerator for vision transformers using dynamic token pruning, ReLU replacement, FFN pruning, and row-wise dataflow to reach 2.31 TOPS/W in 28nm CMOS with under 2% accuracy loss.