SwiftCIM: a 55nm 23.2μJ/Token L-0.5 ReRAM Coupled Digital CIM Accelerator with Fully-Fused Multi-Head Attention Dataflow for FlashAttention

Kunming Shao, Xiaomeng Wang, et al.

ESSERC 2026IEEE European Solid-State Electronics Research Conference (ESSERC)First author

Paper

SwiftCIM overview

A 55nm ReRAM-coupled digital CIM accelerator with a fully fused multi-head attention dataflow for FlashAttention, targeting low energy per token for Transformer inference.

Citation

Kunming Shao, Xiaomeng Wang, and collaborators. SwiftCIM: a 55nm 23.2μJ/Token L-0.5 ReRAM Coupled Digital CIM Accelerator with Fully-Fused Multi-Head Attention Dataflow for FlashAttention. In 2026 IEEE European Solid-State Electronics Research Conference (ESSERC), 2026.