CODA: A Computation-Driven Paradigm for Sparse DNN Acceleration

doi:10.1109/LCA.2025.3637756

CSpace

	CODA: A Computation-Driven Paradigm for Sparse DNN Acceleration
	Liu, Yanhuan 1,2; Li, Wenming 1,2; Zhang, Kunming 1,2; Liu, Tianyu 1,2; Ye, Xiaochun 1,2; An, Xuejun 1,2
	2025-07-01
发表期刊	IEEE COMPUTER ARCHITECTURE LETTERS
ISSN	1556-6056
卷号	24 期号:2 页码:381-384
摘要	The spatial dataflow paradigm, while effective for dense workloads, collapses on unstructured sparsity, manifesting as a profound Twin Crisis: a high Latency Overhead (termed the Latency Crisis) driven by slow, dominant offline software preprocessing for data reordering and formatting, and an Inefficiency Crisis characterized by massive pipeline stalls that leave hardware underutilized. This paper introduces CODA, a novel computation-driven architecture that challenges this data-centric model by shifting the focus from "scheduling data" to "discovering work." CODA materializes this new paradigm through two core innovations. First, a hardware Work Tokenizer eliminates the preprocessing bottleneck on-the-fly, directly resolving the Latency Crisis. Second, a scalable Dynamic Execution Core performs asynchronous, pull-based dispatch of computable work packets to idle resources, which eradicates stalls and solves the Inefficiency Crisis. By directly resolving the Twin Crisis, CODA demonstrates superior performance over state-of-the-art accelerators like Mentor and Hirac. On the BERT-BASE benchmark, it achieves up to 3.95x higher speed, 3.4x greater energy efficiency, and a 3.1x reduction in end-to-end latency, establishing a necessary new architectural blueprint for sparse acceleration.
关键词	Software Hardware Computational modeling Sparse matrices Pipelines Indexes Data models Spatial databases Computational efficiency Vectors Computation-driven architecture sparse DNN acceleration dataflow paradigm unstructured sparsity work tokenizer dynamic execution core asynchronous execution
DOI	10.1109/LCA.2025.3637756
收录类别	SCI
语种	英语
WOS研究方向	Computer Science
WOS类目	Computer Science, Hardware & Architecture
WOS记录号	WOS:001641467300003
出版者	IEEE COMPUTER SOC
引用统计
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/42944
专题	中国科学院计算技术研究所
通讯作者	Li, Wenming
作者单位	1.Chinese Acad Sci, Inst Comp Technol, State Key Lab Processors, Beijing 100045, Peoples R China 2.UCAS, Sch Comp Sci & Technol, Beijing 101408, Peoples R China
推荐引用方式 GB/T 7714	Liu, Yanhuan,Li, Wenming,Zhang, Kunming,et al. CODA: A Computation-Driven Paradigm for Sparse DNN Acceleration[J]. IEEE COMPUTER ARCHITECTURE LETTERS,2025,24(2):381-384.
APA	Liu, Yanhuan,Li, Wenming,Zhang, Kunming,Liu, Tianyu,Ye, Xiaochun,&An, Xuejun.(2025).CODA: A Computation-Driven Paradigm for Sparse DNN Acceleration.IEEE COMPUTER ARCHITECTURE LETTERS,24(2),381-384.
MLA	Liu, Yanhuan,et al."CODA: A Computation-Driven Paradigm for Sparse DNN Acceleration".IEEE COMPUTER ARCHITECTURE LETTERS 24.2(2025):381-384.