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

Benchmark results for ssyrk on Nvidia Geforce Gtx 1650.

Nvidia Geforce Gtx 1650 — wgblas vs cuBLAS

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0108	0.7619	0.0096	0.8533	89.3%
64	0.0147	2.2309	0.0126	2.5990	85.8%
128	0.0244	5.3789	0.0188	6.9660	77.2%
256	0.0840	6.2439	0.0318	16.4829	37.9%
512	0.2969	7.0644	0.1052	19.9319	35.4%
1024	2.1174	3.9618	0.5915	14.1830	27.9%

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

ssyrk-default GB/s chart

ssyrk-default ms chart

See also

ssyrk.js — WebGPU benchmark script
ssyrk.c — CUDA / cuBLAS reference script

Transpose sweep

Unless noted otherwise, every result above uses trans = "no-transpose". trans = "transpose" reads A with a cross-thread lda-strided mirror pattern instead of a coalesced one, and the gap grows with n — collapsed below by default, expand a trans value to see its table and chart.

Nvidia Geforce Gtx 1650 — trans = no-transpose

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
256	0.1314	3.9903	0.0319	16.4498	24.3%
512	0.4362	4.8079	0.1056	19.8504	24.2%
1024	3.1405	2.6711	0.5904	14.2083	18.8%
2048	18.6120	1.8028	3.8093	8.8086	20.5%

ssyrk-transno-transpose GB/s chart

ssyrk-transno-transpose ms chart

Nvidia Geforce Gtx 1650 — trans = transpose

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
256	0.0655	8.0000	0.0352	14.8945	53.7%
512	0.2376	8.8270	0.1088	19.2668	45.8%
1024	1.7384	4.8254	0.6328	13.2557	36.4%
2048	12.2284	2.7440	3.5128	9.5521	28.7%

ssyrk-transtranspose GB/s chart

ssyrk-transtranspose ms chart

See also:

trans.ssyrk.js — WebGPU trans-sweep benchmark script
trans.ssyrk.c — CUDA / cuBLAS trans-sweep reference script

Lda sweep

Unless noted otherwise, every result above uses a tight lda (no padding). Padding the row stride only matters for trans = "transpose" here (swept at both trans values below so that's visible in the data, not just claimed). Collapsed below by default — expand a trans value, then a pad, to see its table and chart.

Nvidia Geforce Gtx 1650 — trans = no-transpose (6 pads)

pad = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.4310	4.8661	0.1043	20.1000	24.2%
1024	2.6357	3.1827	0.5851	14.3377	22.2%
2048	18.6121	1.8028	3.8093	8.8086	20.5%

ssyrk-lda-no-transpose-pad0 GB/s chart

ssyrk-lda-no-transpose-pad0 ms chart

pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.2886	7.2668	0.0962	21.7945	33.3%
1024	1.9596	4.2808	0.5386	15.5741	27.5%
2048	17.7926	1.8859	3.4740	9.6586	19.5%

ssyrk-lda-no-transpose-pad1 GB/s chart

ssyrk-lda-no-transpose-pad1 ms chart

pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.2782	7.5394	0.0956	21.9294	34.4%
1024	1.7039	4.9231	0.4667	17.9748	27.4%
2048	14.0073	2.3955	3.2842	10.2170	23.4%

ssyrk-lda-no-transpose-pad8 GB/s chart

ssyrk-lda-no-transpose-pad8 ms chart

pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.3257	6.4396	0.0900	23.2893	27.7%
1024	1.8938	4.4296	0.4670	17.9618	24.7%
2048	15.3650	2.1838	3.2810	10.2270	21.4%

ssyrk-lda-no-transpose-pad16 GB/s chart

ssyrk-lda-no-transpose-pad16 ms chart

pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.4342	4.8302	0.0898	23.3557	20.7%
1024	2.4775	3.3860	0.4678	17.9323	18.9%
2048	19.1753	1.7499	3.2792	10.2324	17.1%

ssyrk-lda-no-transpose-pad32 GB/s chart

ssyrk-lda-no-transpose-pad32 ms chart

pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.4301	4.8762	0.0900	23.3058	20.9%
1024	2.5067	3.3464	0.4688	17.8932	18.7%
2048	19.4791	1.7226	3.2814	10.2258	16.8%

ssyrk-lda-no-transpose-pad64 GB/s chart

ssyrk-lda-no-transpose-pad64 ms chart

Nvidia Geforce Gtx 1650 — trans = transpose (6 pads)

pad = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.2387	8.7868	0.0938	22.3558	39.3%
1024	1.7433	4.8120	0.5100	16.4498	29.3%
2048	12.2074	2.7487	3.5106	9.5580	28.8%

ssyrk-lda-transpose-pad0 GB/s chart

ssyrk-lda-transpose-pad0 ms chart

pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.2361	8.8814	0.0983	21.3333	41.6%
1024	1.7316	4.8444	0.5200	16.1329	30.0%
2048	12.3521	2.7165	3.6293	9.2454	29.4%

ssyrk-lda-transpose-pad1 GB/s chart

ssyrk-lda-transpose-pad1 ms chart

pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.2356	8.9025	0.0972	21.5792	41.3%
1024	1.6900	4.9635	0.4908	17.0906	29.0%
2048	12.0177	2.7921	3.5547	9.4394	29.6%

ssyrk-lda-transpose-pad8 GB/s chart

ssyrk-lda-transpose-pad8 ms chart

pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.2365	8.8688	0.0963	21.7872	40.7%
1024	1.7144	4.8932	0.4950	16.9470	28.9%
2048	12.1617	2.7590	3.5237	9.5224	29.0%

ssyrk-lda-transpose-pad16 GB/s chart

ssyrk-lda-transpose-pad16 ms chart

pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.2396	8.7521	0.0986	21.2607	41.2%
1024	1.7491	4.7961	0.5070	16.5464	29.0%
2048	12.3984	2.7063	3.5989	9.3236	29.0%

ssyrk-lda-transpose-pad32 GB/s chart

ssyrk-lda-transpose-pad32 ms chart

pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.2396	8.7521	0.0966	21.7007	40.3%
1024	1.7531	4.7850	0.5157	16.2661	29.4%
2048	12.3957	2.7069	3.5958	9.3316	29.0%

ssyrk-lda-transpose-pad64 GB/s chart

ssyrk-lda-transpose-pad64 ms chart

See also:

lda.ssyrk.js — WebGPU lda-sweep benchmark script
lda.ssyrk.c — CUDA / cuBLAS lda-sweep reference script