Unstructured sparse pruning significantly reduces the computational and parametric complexities of deep neural network models. Nevertheless, the highly irregular nature of sparse models limits its performance and effi...
详细信息
Unstructured sparse pruning significantly reduces the computational and parametric complexities of deep neural network models. Nevertheless, the highly irregular nature of sparse models limits its performance and efficiency on traditional computing platforms, thereby prompting the development of specialized hardware solutions. To improve computational efficiency, we introduce the Sparse Dataflow Fusion Accelerator (SPDFA), a specialized architecture meticulously designed for sparse deep neural networks. Firstly, we present a non-blocking data distribution-computing engine that integrates inner product and column product. This engine boosts computational efficiency by decomposing matrix multiplication and convolution into rectangular matrix-vector multiplications. Secondly, we implement a computation array to further exploit the parallelism, and design an on-chip buffer structure that supports multi-line memory access mode. Lastly, to bolster the adaptability of our accelerator, we propose an innovative macroinstruction set coupled with a micro-kernel scheme. Furthermore, we refine the macroinstruction issue strategy, thereby further enhancing computational efficiency. Our evaluation results demonstrate that SPDFA achieves an average 1.29\(\times\)-2.38\(\times\) improvement in computational efficiency compared to the state-of-the-art SpMM accelerators when applied to unstructured sparse deep neural network models. Furthermore, its performance outperforms existing sparse neural network accelerators by a factor of 1.03\(\times\)-1.83\(\times\). Additionally, SPDFA exhibits excellent scalability with a scaling efficiency exceeding 80%.
Welcome to the proceedings of GCC2004 and the city of Wuhan. Grid computing has become a mainstream research area in computer science and the GCC conference has become one of the premier forums for presentation of new...
详细信息
ISBN:
(数字)9783540302087
ISBN:
(纸本)9783540235644
Welcome to the proceedings of GCC2004 and the city of Wuhan. Grid computing has become a mainstream research area in computer science and the GCC conference has become one of the premier forums for presentation of new and exciting research in all aspectsofgridandcooperativecomputing. Theprogramcommitteeispleasedtopresent the proceedings of the 3rd International Conference on Grid and Cooperative Comp- ing (GCC2004), which comprises a collection of excellent technical papers, posters, workshops, and keynote speeches. The papers accepted cover a wide range of exciting topics, including resource grid and service grid, information grid and knowledge grid, grid monitoring,managementand organizationtools, grid portal, grid service, Web s- vices and their QoS, service orchestration, grid middleware and toolkits, software glue technologies, grid security, innovative grid applications, advanced resource reservation andscheduling,performanceevaluationandmodeling,computer-supportedcooperative work, P2P computing, automatic computing, and meta-information management. The conference continues to grow and this year a record total of 581 manuscripts (including workshop submissions) were submitted for consideration. Expecting this growth, the size of the program committee was increased from 50 members for GCC 2003 for 70 in GCC 2004. Relevant differences from previous editions of the conf- ence: it is worth mentioning a signi?cant increase in the number of papers submitted by authors from outside China; and the acceptance rate was much lower than for p- vious GCC conferences. From the 427 papers submitted to the main conference, the program committee selected only 96 regular papers for oral presentation and 62 short papers for poster presentation in the program.
暂无评论