wgblas
    Preparing search index...

    Module benchmarks/nvidia-geforce-gtx-1650/sgemm

    Benchmark results for sgemm on Nvidia Geforce Gtx 1650.

    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0106 1.5515 0.0090 1.8221 85.1%
    64 0.0143 4.5714 0.0101 6.4606 70.8%
    128 0.0237 11.0553 0.0134 19.5513 56.5%
    256 0.0796 13.1651 0.0438 23.9445 55.0%
    512 0.2985 14.0515 0.1998 20.9917 66.9%
    1024 2.0314 8.2588 0.9976 16.8168 49.1%

    Efficiency = wgblas GB/s ÷ cuBLAS GB/s × 100. 100% means parity with cuBLAS; values above 100% mean wgblas achieved greater throughput.

    sgemm-default GB/s chart

    sgemm-default ms chart

    • sgemm.js — WebGPU benchmark script
    • sgemm.c — CUDA / cuBLAS reference script

    Unless noted otherwise, every result above uses transA = transB = "no-transpose". Both shaders load A/B into shared memory with a transpose-dependent index that scatters what would otherwise be a coalesced load — but it's asymmetric: transB dominates (measured +22-57% at n=1024) while transA is small and can even be faster than no-transpose, because B's tile dimension spans a full warp in the coalesced case (so transpose scatters every warp) while A's never gets a full-warp-coalesced load to begin with. All 4 (transA, transB) combinations are swept — collapsed below by default, expand a transA value, then a transB, to see its table and chart (4 combinations total).

    Nvidia Geforce Gtx 1650 — transA = no-transpose (2 transB values)
    transB = no-transpose
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0105 1.5657 0.0083 1.9807 79.1%
    64 0.0143 4.5714 0.0096 6.8495 66.7%
    128 0.0238 11.0256 0.0127 20.6088 53.5%
    256 0.0797 13.1625 0.0432 24.2816 54.2%
    512 0.2990 14.0274 0.1987 21.1134 66.4%
    1024 2.0347 8.2456 0.9964 16.8386 49.0%

    sgemm-trans-no-transpose-no-transpose GB/s chart

    sgemm-trans-no-transpose-no-transpose ms chart

    transB = transpose
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0115 1.4202 0.0086 1.8998 74.8%
    64 0.0150 4.3574 0.0096 6.8040 64.0%
    128 0.0287 9.1429 0.0122 21.4731 42.6%
    256 0.1331 7.8769 0.0330 31.7366 24.8%
    512 0.4396 9.5405 0.1868 22.4573 42.5%
    1024 3.1890 5.2609 0.9844 17.0425 30.9%

    sgemm-trans-no-transpose-transpose GB/s chart

    sgemm-trans-no-transpose-transpose ms chart

    Nvidia Geforce Gtx 1650 — transA = transpose (2 transB values)
    transB = no-transpose
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0107 1.5329 0.0084 1.9431 78.9%
    64 0.0148 4.4329 0.0098 6.7148 66.0%
    128 0.0246 10.6667 0.0136 19.2980 55.3%
    256 0.0856 12.2520 0.0492 21.3333 57.4%
    512 0.3052 13.7421 0.2439 17.1988 79.9%
    1024 2.1678 7.7392 1.0079 16.6451 46.5%

    sgemm-trans-transpose-no-transpose GB/s chart

    sgemm-trans-transpose-no-transpose ms chart

    transB = transpose
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0115 1.4222 0.0084 1.9431 73.2%
    64 0.0156 4.2010 0.0099 6.5958 63.7%
    128 0.0294 8.9043 0.0125 20.8980 42.6%
    256 0.1396 7.5096 0.0562 18.6553 40.3%
    512 0.4631 9.0576 0.2336 17.9588 50.4%
    1024 3.3841 4.9576 0.9873 16.9934 29.2%

    sgemm-trans-transpose-transpose GB/s chart

    sgemm-trans-transpose-transpose ms chart

    See also:

    Unless noted otherwise, every result above uses a tight lda/ldb/ldc (no padding). lda and ldc were scoped and found to be non-effects; padding ldb only matters for transB = "transpose" here (swept at both transB values below so that's visible in the data). Collapsed below by default — expand a transB value, then a pad, to see its table and chart.

    Nvidia Geforce Gtx 1650 — transB = no-transpose (6 pads)
    pad = 0
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0103 1.5876 0.0086 1.9033 83.4%
    64 0.0143 4.5714 0.0097 6.7479 67.7%
    128 0.0233 11.2605 0.0134 19.5981 57.5%
    256 0.0795 13.1969 0.0436 24.0587 54.9%
    512 0.2970 14.1241 0.1986 21.1219 66.9%
    1024 2.0308 8.2613 0.9967 16.8327 49.1%

    sgemm-ldb-no-transpose-pad0 GB/s chart

    sgemm-ldb-no-transpose-pad0 ms chart

    pad = 1
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0105 1.5657 0.0083 1.9807 79.1%
    64 0.0143 4.5714 0.0097 6.7702 67.5%
    128 0.0238 11.0182 0.0136 19.2980 57.1%
    256 0.0795 13.1969 0.0444 23.6251 55.9%
    512 0.2985 14.0530 0.2009 20.8780 67.3%
    1024 2.0350 8.2445 1.0002 16.7735 49.2%

    sgemm-ldb-no-transpose-pad1 GB/s chart

    sgemm-ldb-no-transpose-pad1 ms chart

    pad = 8
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0105 1.5657 0.0085 1.9357 80.9%
    64 0.0143 4.5714 0.0100 6.5746 69.5%
    128 0.0236 11.1153 0.0130 20.2022 55.0%
    256 0.0794 13.1996 0.0431 24.3357 54.2%
    512 0.2970 14.1241 0.1923 21.8090 64.8%
    1024 2.0316 8.2581 0.9891 16.9626 48.7%

    sgemm-ldb-no-transpose-pad8 GB/s chart

    sgemm-ldb-no-transpose-pad8 ms chart

    pad = 16
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0105 1.5586 0.0087 1.8928 82.3%
    64 0.0143 4.5714 0.0100 6.5641 69.6%
    128 0.0235 11.1380 0.0126 20.8183 53.5%
    256 0.0792 13.2369 0.0430 24.3810 54.3%
    512 0.2967 14.1356 0.1950 21.5137 65.7%
    1024 2.0383 8.2309 0.9892 16.9609 48.5%

    sgemm-ldb-no-transpose-pad16 GB/s chart

    sgemm-ldb-no-transpose-pad16 ms chart

    pad = 32
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0105 1.5586 0.0083 1.9692 79.1%
    64 0.0143 4.5714 0.0096 6.8495 66.7%
    128 0.0236 11.1304 0.0125 20.9782 53.1%
    256 0.0792 13.2369 0.0432 24.2726 54.5%
    512 0.2978 14.0862 0.1987 21.1134 66.7%
    1024 2.0366 8.2378 0.9942 16.8752 48.8%

    sgemm-ldb-no-transpose-pad32 GB/s chart

    sgemm-ldb-no-transpose-pad32 ms chart

    pad = 64
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0113 1.4484 0.0083 1.9730 73.4%
    64 0.0143 4.5714 0.0097 6.7368 67.9%
    128 0.0235 11.1532 0.0125 20.8980 53.4%
    256 0.0792 13.2316 0.0430 24.3810 54.3%
    512 0.2977 14.0900 0.1984 21.1372 66.7%
    1024 2.0375 8.2344 0.9950 16.8622 48.8%

    sgemm-ldb-no-transpose-pad64 GB/s chart

    sgemm-ldb-no-transpose-pad64 ms chart

    Nvidia Geforce Gtx 1650 — transB = transpose (6 pads)
    pad = 0
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0111 1.4734 0.0082 2.0000 73.7%
    64 0.0159 4.1207 0.0098 6.6928 61.6%
    128 0.0287 9.1480 0.0124 21.1953 43.2%
    256 0.1332 7.8694 0.0329 31.8290 24.7%
    512 0.4389 9.5561 0.1864 22.5055 42.5%
    1024 3.1770 5.2808 0.9851 17.0317 31.0%

    sgemm-ldb-transpose-pad0 GB/s chart

    sgemm-ldb-transpose-pad0 ms chart

    pad = 1
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0100 1.6463 0.0084 1.9542 84.2%
    64 0.0123 5.3333 0.0098 6.6819 79.8%
    128 0.0208 12.6031 0.0123 21.3333 59.1%
    256 0.0771 13.5939 0.0341 30.7536 44.2%
    512 0.2668 15.7236 0.1864 22.5055 69.9%
    1024 1.9606 8.5572 0.9851 17.0309 50.2%

    sgemm-ldb-transpose-pad1 GB/s chart

    sgemm-ldb-transpose-pad1 ms chart

    pad = 8
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0101 1.6254 0.0083 1.9655 82.7%
    64 0.0123 5.3333 0.0099 6.6494 80.2%
    128 0.0219 11.9766 0.0123 21.3333 56.1%
    256 0.0906 11.5768 0.0336 31.1779 37.1%
    512 0.2499 16.7869 0.1867 22.4650 74.7%
    1024 1.8726 8.9592 0.9835 17.0583 52.5%

    sgemm-ldb-transpose-pad8 GB/s chart

    sgemm-ldb-transpose-pad8 ms chart

    pad = 16
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0100 1.6358 0.0084 1.9468 84.0%
    64 0.0128 5.1393 0.0096 6.8040 75.5%
    128 0.0228 11.5137 0.0122 21.4731 53.6%
    256 0.1018 10.3012 0.0331 31.6599 32.5%
    512 0.2888 14.5248 0.1864 22.5055 64.5%
    1024 2.1069 7.9629 0.9836 17.0575 46.7%

    sgemm-ldb-transpose-pad16 GB/s chart

    sgemm-ldb-transpose-pad16 ms chart

    pad = 32
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0102 1.6101 0.0082 1.9961 80.7%
    64 0.0136 4.8245 0.0102 6.4000 75.4%
    128 0.0246 10.6667 0.0122 21.4731 49.7%
    256 0.1132 9.2630 0.0341 30.7392 30.1%
    512 0.3740 11.2142 0.1864 22.5036 49.8%
    1024 2.7133 6.1833 0.9851 17.0312 36.3%

    sgemm-ldb-transpose-pad32 GB/s chart

    sgemm-ldb-transpose-pad32 ms chart

    pad = 64
    n wgblas ms wgblas GB/s cuBLAS ms cuBLAS GB/s efficiency
    32 0.0102 1.6000 0.0083 1.9655 81.4%
    64 0.0135 4.8473 0.0098 6.6928 72.4%
    128 0.0246 10.6667 0.0121 21.7294 49.1%
    256 0.1147 9.1429 0.0337 31.1187 29.4%
    512 0.3735 11.2301 0.1866 22.4804 50.0%
    1024 2.7383 6.1269 0.9846 17.0403 36.0%

    sgemm-ldb-transpose-pad64 GB/s chart

    sgemm-ldb-transpose-pad64 ms chart

    See also:

    Settings
    Member Visibility

    On This Page
    Nvidia Geforce Gtx 1650 — wgblas vs cuBLASSee alsoTranspose sweepLdb sweep