Onyx：稀疏张量代数的可编程加速器.pdf

《Onyx：稀疏张量代数的可编程加速器.pdf》由会员分享，可在线阅读，更多相关《Onyx：稀疏张量代数的可编程加速器.pdf（91页珍藏版）》请在三个皮匠报告上搜索。

1、Onyx:A Programmable Accelerator for Sparse Tensor AlgebraKalhan Koul1,Maxwell Strange1,Jackson Melchert1,Alex Carsello1,Yuchen Mei1,Olivia Hsu1,Taeyoung Kong1,Po-Han Chen1,Huifeng Ke1,Keyi Zhang1,Qiaoyi Liu1,Gedeon Nyengele1,Akhilesh Balasingam1,Jayashree Adivarahan1,Ritvik Sharma1,Zhouhua Xie1,Chri

2、stopher Torng2,Joel Emer3,Fredrik Kjolstad1,Mark Horowitz1,Priyanka Raina11Stanford University,CA,USA;2University of Southern California,CA,USA;3MIT,MA,USASparse Applications2Applications ranging from scientific computing to machine learning can have extremelyextremely sparse inputs Sparse Applicati

3、onsApplications ranging from scientific computing to machine learning can have extremelyextremely sparse inputs Scientific ComputingUp to 99.9%sparse matrices3Sparse ApplicationsApplications ranging from scientific computing to machine learning can have extremelyextremely sparse inputs Scientific Co

4、mputingGraph Neural NetworksLayer1Layer2Up to 99.9%sparse matricesUp to 99.9%sparse matrices4Sparse ApplicationsApplications ranging from scientific computing to machine learning can have extremelyextremely sparse inputs Scientific ComputingUp to 99.9%sparse matricesGraph Neural NetworksLayer1Layer2

5、Up to 99.9%sparse matricesSparse TransformersUp to 93%sparse matrices5Hardware accelerators for end-to-end applications exploiting sparsity are fixed function and become obsolete as applications evolveEnd-to-End Sparse AcceleratorsTransformer InferenceCNN+Graph Conv Network 2D/3D CNN for Point Cloud

6、sSTPISSCC 2023AR SoCVLSI 2022SCONV ISSCC 2023GPPU VLSI 20233D Navigation6Hardware accelerators for end-to-end applications exploiting sparsity are fixed function and become obsolete as applications evolveEnd-to-End Sparse AcceleratorsTransformer InferenceCNN+Graph Conv Network 2D/3D CNN for Point Cl