Datasets (Benchmarks)

Large Scale Matrix Multiplication

• All the matrices are generated by the Matrix Manager library [git repository]

  • You can read the matrix by the library too

• Related papers

  • HPMaX: Heterogeneous Parallel Matrix Multiplication using CPUs and GPUs [link]

• (Dataset) Single-precision floating point numbers (the ratio of zero)

  • 8192 by 8192 (0.25 GB) : TypeA(10%, 40%, 70%), TypeB(10%,40%, 70%)

  • 16384 by 16384 (1 GB) : TypeA(10%, 40%, 70%), TypeB(10%, 40%, 70%)

  • 32768 by 16384 (2 GB) : TypeA(10%, 40%, 70%) / 16384 by 32768 (2GB) : TypeB(10%, 40%, 70%)

  • 32768 by 32768 (4 GB) : TypeA(10%, 40%, 70%), TypeB(10%, 40%, 70%)

  • 65536 by 32768, 32768 by 65536, 65536 by 65536,131072 by 65536, 65536 by 131072, 131072 by 131072

    • contact us (hpchomin at gmail dot com)

• (Dataset) Double-precision floating point numbers (the ratio of zero)

  • 8192 by 8192 (0.5 GB) : TypeA(10%, 40%, 70%), TypeB(10%, 40%, 70%)

  • 16384 by 16384 (2 GB) : TypeA(10%, 40%, 70%), TypeB(10%, 40%, 70%)

  • 32768 by 16384 (4 GB) : TypeA(10%, 40%, 70%) / 16384 by 32768 (4 GB) : TypeB(10%, 40%, 70%)

  • 32768 by 32768 (8 GB) : TypeA(10%,40%, 70%), TypeB(10%, 40%, 70%)

  • 65536 by 32768, 32768 by 65536, 65536 by 65536, 131072 by 65536, 65536 by 131072, 131072 by 131072

    • contact us (hpchomin at gmail dot com)

• Data Construction program

  • If you want to make different number of matrix size, use this program.

  • It automatically makes matrices of 10%, 40%, 70% (the ratio of zeros) with single and double precision.

  • CMD command

    • dataconstruction.exe (matrix row) (matrix col)

    • When it finish each work, it prints the message like (matrixrow)(ratio)per.bin