Publication:
Flexagon: a multi-dataflow sparse-sparse matrix multiplication accelerator for efficient DNN processing

Simple item page
dc.contributor.authorGarg, Raveesh
dc.contributor.authorPellauer, Michael
dc.contributor.authorKrishna, Tushar
dc.contributor.authorMuñoz Martínez, Francisco
dc.contributor.authorAbellán Miguel, José Luis
dc.contributor.authorAcacio Sánchez, Manuel Eugenio
dc.contributor.departmentIngeniería y Tecnología de Computadores
dc.contributor.otherFacultad de Informática
dc.date.accessioned2023-02-17T11:30:48Z
dc.date.available2023-02-17T11:30:48Z
dc.date.copyright© 2023 ACM
dc.date.issued2023-03-25
dc.description.abstractSparsity is a growing trend in modern DNN models.Existing Sparse-Sparse Matrix Multiplication (SpMSpM) accel-erators are tailored to a particular SpMSpM dataflow (i.e., InnerProduct, Outer Product or Gustavson’s), which determines theiroverall efficiency. We demonstrate that this static decision inher-ently results in a suboptimal dynamic solution. This is becausedifferent SpMSpM kernels show varying features (i.e., dimensions,sparsity pattern, sparsity degree), which makes each dataflow bettersuited to different data sets.In this work we present Flexagon, the first SpMSpM reconfig-urable accelerator that is capable of performing SpMSpM computa-tion by using the particular dataflow that best matches each case.Flexagon accelerator is based on a novel Merger-Reduction Net-work (MRN) that unifies the concept of reducing and merging inthe same substrate, increasing efficiency. Additionally, Flexagonalso includes a new L1 on-chip memory organization, specificallytailored to the different access characteristics of the input and out-put compressed matrices. Using detailed cycle-level simulation ofcontemporary DNN models from a variety of application domains,we show that Flexagon achieves average performance benefits of4.59×, 1.71×, and 1.35×with respect to the state-of-the-art SIGMA-like, SpArch-like and GAMMA-like accelerators (265%, 67%, and18%, respectively, in terms of average performance/area efficiency).es
dc.formatapplication/pdfes
dc.format.extent14es
dc.identifier.citationASPLOS 2023: Proceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3. pp. 252-265
dc.identifier.doihttps://doi.org/10.1145/3582016.3582069
dc.identifier.eissn978-1-4503-9918-0/23/03.
dc.identifier.urihttp://hdl.handle.net/10201/128556
dc.languageenges
dc.publisherAssociation for Computing Machinery
dc.relationThis work was supported by grants TED2021-130233B-C33 and RYC2021-031966-I both funded by MCIN/AEI/10.13039/501100011033 and by the European Union NextGenerationEU/PRTR. F. Muñoz-Martínez was supported by grant 20749/FPI/18 from Fundación Séneca. A part of the work was supported by the ARIAA co-design center funded by the U.S. Department of Energy (DOE) Office of Science, Advanced Scientific Computing Research program. es
dc.relation.ispartofASPLOS ’23, March 25–29, 2023, Vancouver, BC, Canadaes
dc.relation.publisherversionhttps://dl.acm.org/doi/10.1145/3582016.3582069
dc.rightsinfo:eu-repo/semantics/openAccesses
dc.rightsAttribution-NonCommercial-NoDerivatives 4.0 Internacional*
dc.rights.urihttp://creativecommons.org/licenses/by-nc-nd/4.0/*
dc.subjectDeep Neural Network Acceleratorses
dc.subjectSparse-Sparse Matrix Multiplicationes
dc.subjectDataflowes
dc.subjectMerger-Reduction Networkes
dc.subjectMemory Hierarchyes
dc.titleFlexagon: a multi-dataflow sparse-sparse matrix multiplication accelerator for efficient DNN processinges
dc.typeinfo:eu-repo/semantics/articlees
dc.type.versioninfo:eu-repo/semantics/AcceptedVersión
dspace.entity.typePublicationes
relation.isAuthorOfPublicationdbf4e7f7-682b-4a97-a941-16178079ac7a
relation.isAuthorOfPublication0408fa9f-b61e-48cb-a08f-ae8999f885ee
relation.isAuthorOfPublication5aeaee45-9977-413f-94ee-2a27e25672f8
relation.isAuthorOfPublication.latestForDiscoverydbf4e7f7-682b-4a97-a941-16178079ac7a
Files
Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
asplosc23main-p1166-p-f253962830-63228-submitted.pdf
Size:
1.47 MB
Format:
Adobe Portable Document Format
Description:
Download
License bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
license.txt
Size:
1.39 KB
Format:
Item-specific license agreed upon to submission
Description:
Download
Collections
Artículos