6461 - Challenges and Trends in Sparse Matrix Multiplication on HPC Workloads.pdf

1、Hector Arroyo Gonzalez|OPENCHIPChallenges and Trends in Sparse Matrix Multiplication on HPC WorkloadsChallenges and Trends in Sparse Matrix Multiplication on HPC WorkloadsHector Arroyo Gonzalez Research Engineer|Innovation Dpt|OPENCHIPAck-Erich FochtARTIFICIAL INTELLIGENCE(AI)Introduce the challenge

2、 of sparse data workloads and why is it relevantCover the existing methods to represent sparse dataExplore some of the hardware architecture design trends w/r sparsityDetail the key steps for sparse calculationGo in detail into how current research approaches solve this challengeBriefly Introduce a

3、proprietary line of researchKey insights from the talkDiscussion/Questions?In todays presentationWhat qualifies as a sparse matrix?A brief introduction to the sparsity challengeThe challenge“a matrix whose number of non-zero entries(NNZ)is().”2How does sparse HPC workloads look like?In many HPC work

4、loads,data is represented as large matrices that are predominantly filled with zeros.This sparsity poses unique computational and storage challenges,requiring specialized techniques to handle them efficiently.“any matrix with enough zeros that it pays to take advantage of them.”1Key Observations-Lev

5、eraging sparsity drastically reduces the total number of operations.-Yields potential multi-order-of-magnitude performance improvements.-Improves energy efficiency by focusing on nonzero elements only.Implementation Requirements-Requires a fundamental reevaluation of both software and hardware archi

6、tectures to efficiently exploit sparse data structures.Amazon product co-purchasing network 3Sparsity:0.974128Sparsity:0.999982Profiling of single cells-PBMC3k Gene Expression Matrix 4 Specific workflow for sparse matrix calculationTraditional vs Sparsity-adapted workflowTraditional workflowSparse I