BROWSE
Detail
Journal Article
An Operand Divergence-Aware All-Bank PIM Architecture for Sparse Attention
Cited 0 time in 
Share 
Share
- Authors
- Soojin Hwang, Sanghyeon Lee, Juhyun Lee, Jaehyuk Huh
- Issue Date
- 2026-04
- Citation
- IEEE Computer Architecture Letters, v.25, pp.158-161
- ISSN
- 1556-6056
- Publisher
- IEEE
- Language
- English
- Type
- Journal Article
- Abstract
- Processing-in-memory (PIM) architectures leverage all-bank execution to provide high internal bandwidth for memory-bound operations. However, synchronized control across banks fundamentally conflicts with the irregular access patterns of sparse workloads, where operand divergence requires banks to access different row or column addresses. We propose a novel divergence-aware all-bank PIM architecture that resolves operand-level data divergence. Our design enables each bank to dynamically access distinct rows or columns under a shared command, allowing sparse execution without sacrificing bank-level parallelism. We apply this architecture to accelerate large language models with dynamic KV-cache filtering with unstructured token-level sparsity. The simulation results show that the proposed design achieves an average 6.6× speedup over dense all-bank PIM.
- Keyword
- Processing-in-memory (PIM), data divergence