A 28nm 0.22{\mu}J/token memory-compute-intensity-aware CNN-Transformer accelerator with hybrid-attention-based layer-fusion and cascaded pruning for semantic-segmentation

TL;DR

This paper introduces a 28nm CNN-Transformer accelerator optimized for semantic segmentation, combining hybrid attention, layer fusion, and cascaded pruning to significantly improve energy efficiency and reduce power consumption.

Contribution

It presents a novel 28nm accelerator with hybrid attention, layer-fusion, and cascaded pruning, achieving substantial energy savings for semantic segmentation tasks.

Findings

Achieves 3.86 to 10.91 times energy reduction compared to prior designs.

Attains peak energy efficiency of 52.90 TOPS/W (INT8).

Features hybrid attention unit, layer-fusion scheduler, and cascaded feature-map pruner.

Abstract

This work presents a 28nm 13.93mm2 CNN-Transformer accelerator for semantic segmentation, achieving 3.86-to-10.91x energy reduction over previous designs. It features a hybrid attention unit, layer-fusion scheduler, and cascaded feature-map pruner, with peak energy efficiency of 52.90TOPS/W (INT8).