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

Benchmark results for dsymv 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	0.0082	0.6094	0.0039	1.2946	47.1%
64	0.0085	2.1434	0.0042	4.2868	50.0%
128	0.0115	6.0167	0.0054	12.7434	47.2%
256	0.0228	11.8285	0.0073	37.0749	31.9%
512	0.0561	18.9427	0.0161	66.1015	28.7%
1024	0.1779	23.7353	0.0598	70.6089	33.6%
1280	0.2823	23.3443	0.0820	80.3277	29.1%
2048	0.6369	26.4309	0.1963	85.7716	30.8%
4096	3.5168	19.1150	0.7721	87.0663	22.0%

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

See also

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

Stride sweep

Unless noted otherwise, every result above uses unit stride (incx = incy = 1) — the normal case, and the coalesced, best-case GPU access pattern. Real usage sometimes passes a non-unit stride (e.g. operating on a row or column of a larger matrix, where incx = lda), which breaks memory coalescing and costs measurably more. This section sweeps a few representative strides to characterize that cost separately, collapsed below by default — expand a stride to see its table and chart.

Nvidia Geforce Gtx 1650 — stride = 4

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5748	0.0039	1.2033	47.8%
64	0.0086	2.0520	0.0049	3.6197	56.7%
128	0.0119	5.7358	0.0055	12.3362	46.5%
256	0.0232	11.5359	0.0075	35.6923	32.3%
512	0.0594	17.8276	0.0163	65.0059	27.4%
1024	0.1946	21.6632	0.0613	68.7791	31.5%
1280	0.3107	21.1773	0.0841	78.2493	27.1%
2048	0.7258	23.1722	0.2007	83.7959	27.7%
4096	3.3814	19.8708	0.7841	85.6955	23.2%

Nvidia Geforce Gtx 1650 — stride = 5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0084	0.5670	0.0039	1.2131	46.7%
64	0.0088	2.0073	0.0044	4.0588	49.5%
128	0.0120	5.6671	0.0060	11.3191	50.1%
256	0.0225	11.8636	0.0074	36.3130	32.7%
512	0.0572	18.5263	0.0160	66.0439	28.1%
1024	0.1925	21.8973	0.0613	68.7791	31.8%
1280	0.3154	20.8604	0.0840	78.3537	26.6%
2048	0.7594	22.1455	0.2023	83.1463	26.6%
4096	3.7059	18.1307	0.8342	80.5429	22.5%

Nvidia Geforce Gtx 1650 — stride = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0084	0.5649	0.0041	1.1562	48.9%
64	0.0095	1.8617	0.0047	3.7297	49.9%
128	0.0145	4.7028	0.0059	11.5027	40.9%
256	0.0329	8.1285	0.0082	32.6250	24.9%
512	0.0985	10.7481	0.0183	57.7452	18.6%
1024	0.3482	12.1059	0.0676	62.3636	19.4%
1280	0.5489	11.9869	0.0920	71.5504	16.8%
2048	1.3073	12.8652	0.2048	82.1200	15.7%
4096	8.7356	7.6916	0.7940	84.6214	9.1%

Nvidia Geforce Gtx 1650 — stride = 33

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0083	0.5692	0.0040	1.1840	48.1%
64	0.0087	2.0220	0.0042	4.1818	48.4%
128	0.0123	5.5417	0.0054	12.6667	43.7%
256	0.0242	11.0257	0.0072	36.8742	29.9%
512	0.0707	14.9719	0.0174	60.9355	24.6%
1024	0.2642	15.9516	0.0632	66.7234	23.9%
1280	0.4321	15.2251	0.0866	75.9512	20.0%
2048	1.0495	16.0244	0.2040	82.4485	19.4%
4096	7.9527	8.4488	0.7939	84.6368	10.0%

Nvidia Geforce Gtx 1650 — stride = 255

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0084	0.5670	0.0040	1.1746	48.3%
64	0.0089	1.9856	0.0042	4.1977	47.3%
128	0.0121	5.6148	0.0053	12.8193	43.8%
256	0.0240	11.1360	0.0072	36.9558	30.1%
512	0.0716	14.7979	0.0164	64.4990	22.9%
1024	0.2683	15.7099	0.0630	66.8928	23.5%
1280	0.4383	15.0117	0.0861	76.4454	19.6%
2048	1.0673	15.7580	0.2048	82.1072	19.2%
4096	16.4628	4.0814	0.7976	84.2361	4.8%

Nvidia Geforce Gtx 1650 — stride = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0083	0.5714	0.0040	1.1746	48.6%
64	0.0096	1.8492	0.0047	3.7679	49.1%
128	0.0144	4.7289	0.0054	12.7045	37.2%
256	0.0328	8.1563	0.0081	33.0119	24.7%
512	0.0984	10.7656	0.0164	64.6250	16.7%
1024	0.3502	12.0351	0.0620	67.9278	17.7%
1280	0.5492	11.9786	0.0864	76.1199	15.7%
2048	1.3118	12.8211	0.2048	82.1328	15.6%
4096	12.7392	5.2744	0.7982	84.1804	6.3%

See also:

• stride.dsymv.js — WebGPU stride-sweep benchmark script
• stride.dsymv.c — CUDA / cuBLAS stride-sweep reference script

Uplo sweep

Unless noted otherwise, every result above uses uplo = "lower". Real workgroups dispatch in increasing index order, so uplo = "upper" front-loads the heaviest rows first (worse — long-running heavy workgroups have nothing to overlap with) while lower back-loads them (better — light rows clear fast, the heavy tail gets full GPU to itself) — collapsed below by default, expand a uplo value to see its table and chart.

Nvidia Geforce Gtx 1650 — uplo = lower

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5759	0.0038	1.2542	45.9%
64	0.0085	2.0791	0.0044	3.9712	52.4%
128	0.0114	5.9691	0.0058	11.6923	51.1%
256	0.0220	12.1307	0.0072	37.2027	32.6%
512	0.0553	19.1481	0.0156	67.8033	28.2%
1024	0.1781	23.6615	0.0597	70.6420	33.5%
1280	0.2849	23.0946	0.0822	80.0467	28.9%
2048	0.6466	26.0118	0.1964	85.6392	30.4%
4096	2.9778	22.5642	0.7719	87.0473	25.9%

Nvidia Geforce Gtx 1650 — uplo = upper

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0083	0.5714	0.0039	1.2231	46.7%
64	0.0086	2.0520	0.0043	4.1504	49.4%
128	0.0118	5.7748	0.0053	12.8193	45.0%
256	0.0228	11.7139	0.0078	34.3704	34.1%
512	0.0558	18.9725	0.0162	65.1980	29.1%
1024	0.1777	23.7127	0.0616	68.3863	34.7%
1280	0.2820	23.3345	0.0840	78.3537	29.8%
2048	0.6373	26.3886	0.2022	83.1858	31.7%
4096	3.4328	19.5732	0.7952	84.4921	23.2%

See also:

• uplo.dsymv.js — WebGPU uplo-sweep benchmark script
• uplo.dsymv.c — CUDA / cuBLAS uplo-sweep reference script

Lda sweep

Unless noted otherwise, every result above uses a tight lda (no padding). Padding the row stride changes throughput here — the exact mechanism and shape of that effect is routine-specific — collapsed below by default, expand a pad value to see its table and chart.

Nvidia Geforce Gtx 1650 — pad = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.6094	0.0039	1.2683	48.0%
64	0.0085	2.1273	0.0043	4.2388	50.2%
128	0.0116	5.9586	0.0053	12.9341	46.1%
256	0.0220	12.2504	0.0077	35.1399	34.9%
512	0.0553	19.2222	0.0161	66.1673	29.1%
1024	0.1784	23.6778	0.0595	71.0078	33.3%
1280	0.2849	23.1280	0.0822	80.1869	28.8%
2048	0.6460	26.0603	0.1964	85.7227	30.4%
4096	2.9383	22.8786	0.7722	87.0519	26.3%

Nvidia Geforce Gtx 1650 — pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0083	0.6000	0.0041	1.2188	49.2%
64	0.0085	2.1353	0.0045	4.0571	52.6%
128	0.0111	6.2248	0.0054	12.7059	49.0%
256	0.0184	14.6111	0.0075	36.0428	40.5%
512	0.0430	24.7143	0.0164	64.6855	38.2%
1024	0.1346	31.3687	0.0599	70.4957	44.5%
1280	0.2355	27.9783	0.0839	78.5205	35.6%
2048	0.5038	33.4146	0.1983	84.9064	39.4%

Nvidia Geforce Gtx 1650 — pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.6070	0.0039	1.2787	47.5%
64	0.0085	2.1353	0.0044	4.1765	51.1%
128	0.0114	6.0589	0.0055	12.5217	48.4%
256	0.0193	13.9685	0.0075	36.0428	38.8%
512	0.0467	22.7819	0.0164	64.8750	35.1%
1024	0.1454	29.0518	0.0594	71.1034	40.9%
1280	0.2346	28.0870	0.0831	79.2610	35.4%
2048	0.5335	31.5547	0.1982	84.9270	37.2%

Nvidia Geforce Gtx 1650 — pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0083	0.6047	0.0039	1.2787	47.3%
64	0.0085	2.1273	0.0045	4.0427	52.6%
128	0.0122	5.6842	0.0055	12.5948	45.1%
256	0.0225	11.9545	0.0072	37.1567	32.2%
512	0.0696	15.2647	0.0163	65.2574	23.4%
1024	0.2205	19.1480	0.0594	71.1034	26.9%
1280	0.3506	18.7952	0.0822	80.1557	23.4%
2048	0.8847	19.0278	0.1973	85.3264	22.3%

Nvidia Geforce Gtx 1650 — pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0086	0.5821	0.0039	1.2735	45.7%
64	0.0090	2.0214	0.0043	4.2547	47.5%
128	0.0123	5.6250	0.0054	12.7434	44.1%
256	0.0233	11.5684	0.0074	36.5913	31.6%
512	0.0580	18.3210	0.0162	65.6443	27.9%
1024	0.1925	21.9362	0.0593	71.1802	30.8%
1280	0.2833	23.2599	0.0822	80.1713	29.0%
2048	0.6869	24.5076	0.1966	85.6250	28.6%

Nvidia Geforce Gtx 1650 — pad = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0085	0.5843	0.0039	1.2893	45.3%
64	0.0087	2.0882	0.0043	4.2547	49.1%
128	0.0120	5.7677	0.0053	13.0909	44.1%
256	0.0222	12.1443	0.0073	36.8315	33.0%
512	0.0562	18.9265	0.0162	65.7743	28.8%
1024	0.1823	23.1685	0.0608	69.4934	33.3%
1280	0.2703	24.3750	0.0827	79.6904	30.6%
2048	0.6807	24.7299	0.1966	85.6250	28.9%

See also:

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

alpha sweep

alpha is a plain multiplier here: the kernel applies it unconditionally, with no branch for any particular value. A flat sweep is therefore the expected result and is recorded as a measured null. Levels include 0, 1 and a denormal-producing 1e-38 because those are the values a shader could special-case if it ever grew a branch — and strsm is the routine where one does.

Nvidia Geforce Gtx 1650 — alpha = -3.75

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0083	0.5736	0.0041	1.1562	49.6%
64	0.0085	2.0674	0.0043	4.1194	50.2%
128	0.0114	5.9608	0.0061	11.0833	53.8%
256	0.0220	12.1572	0.0072	37.2857	32.6%
512	0.0553	19.1481	0.0156	67.6646	28.3%
1024	0.1785	23.6128	0.0595	70.8129	33.3%
1280	0.2852	23.0648	0.0819	80.3125	28.7%
2048	0.6452	26.0666	0.1962	85.7091	30.4%
4096	2.9411	22.8457	0.7721	87.0239	26.3%

Nvidia Geforce Gtx 1650 — alpha = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0102	0.4625	0.0045	1.0571	43.8%
64	0.0107	1.6453	0.0054	3.2471	50.7%
128	0.0150	4.5277	0.0064	10.5608	42.9%
256	0.0274	9.7684	0.0088	30.4818	32.0%
512	0.0704	15.0468	0.0200	52.8140	28.5%
1024	0.2270	18.5667	0.0722	58.3442	31.8%
1280	0.3552	18.5242	0.0963	68.3511	27.1%
2048	0.8190	20.5348	0.2274	73.9664	27.8%
4096	2.9711	22.6145	0.8744	76.8393	29.4%

Nvidia Geforce Gtx 1650 — alpha = 1e-38

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0083	0.5725	0.0038	1.2333	46.4%
64	0.0086	2.0597	0.0043	4.1504	49.6%
128	0.0115	5.9276	0.0053	12.8193	46.2%
256	0.0220	12.1219	0.0071	37.6216	32.2%
512	0.0553	19.1481	0.0160	66.1099	29.0%
1024	0.1782	23.6552	0.0595	70.8510	33.4%
1280	0.2853	23.0584	0.0823	79.9689	28.8%
2048	0.6456	26.0492	0.1965	85.5974	30.4%
4096	2.9972	22.4175	0.7728	86.9464	25.8%

Nvidia Geforce Gtx 1650 — alpha = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5759	0.0048	0.9834	58.6%
64	0.0085	2.0674	0.0057	3.1186	66.3%
128	0.0116	5.8866	0.0064	10.6135	55.5%
256	0.0221	12.0868	0.0088	30.2609	39.9%
512	0.0552	19.1926	0.0198	53.4540	35.9%
1024	0.1782	23.6552	0.0717	58.8000	40.2%
1280	0.2847	23.1063	0.0959	68.6019	33.7%
2048	0.6472	25.9873	0.1966	85.5347	30.4%
4096	2.9861	22.5013	0.7721	87.0239	25.9%

Nvidia Geforce Gtx 1650 — alpha = 2.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5770	0.0038	1.2333	46.8%
64	0.0085	2.0674	0.0043	4.1504	49.8%
128	0.0119	5.7128	0.0054	12.6667	45.1%
256	0.0220	12.1219	0.0072	36.8742	32.9%
512	0.0553	19.1481	0.0156	67.8033	28.2%
1024	0.1801	23.4030	0.0594	70.9655	33.0%
1280	0.2853	23.0635	0.0819	80.3125	28.7%
2048	0.6474	25.9764	0.1959	85.8561	30.3%
4096	2.9226	22.9898	0.7723	86.9968	26.4%

See also:

• alpha.dsymv.js — WebGPU alpha-sweep benchmark script
• alpha.dsymv.c — CUDA / cuBLAS alpha-sweep reference script

beta sweep

beta scales the existing y/C before accumulation. Reference BLAS is permitted to skip reading that operand entirely when beta is 0, so unlike alpha this sweep has a mechanism to be non-flat — a step at 0 means the shortcut is taken, and its size is what it saves.

Nvidia Geforce Gtx 1650 — beta = -3.75

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0083	0.5692	0.0040	1.1746	48.5%
64	0.0086	2.0559	0.0044	4.0292	51.0%
128	0.0117	5.8142	0.0055	12.3006	47.3%
256	0.0221	12.0956	0.0072	36.9558	32.7%
512	0.0553	19.1481	0.0161	65.7813	29.1%
1024	0.1782	23.6552	0.0601	70.1716	33.7%
1280	0.2850	23.0882	0.0827	79.5666	29.0%
2048	0.6464	26.0176	0.1971	85.3126	30.5%
4096	2.9926	22.4525	0.7719	87.0491	25.8%

Nvidia Geforce Gtx 1650 — beta = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5759	0.0038	1.2437	46.3%
64	0.0086	2.0520	0.0042	4.1818	49.1%
128	0.0115	5.9111	0.0056	12.1254	48.7%
256	0.0220	12.1572	0.0071	37.4529	32.5%
512	0.0553	19.1481	0.0158	67.1156	28.5%
1024	0.1782	23.6552	0.0594	70.9273	33.4%
1280	0.2848	23.0985	0.0821	80.1091	28.8%
2048	0.6453	26.0621	0.1969	85.4235	30.5%
4096	2.9770	22.5699	0.7729	86.9302	26.0%

Nvidia Geforce Gtx 1650 — beta = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5759	0.0040	1.1840	48.6%
64	0.0086	2.0636	0.0044	4.0588	50.8%
128	0.0116	5.8703	0.0057	11.9887	49.0%
256	0.0220	12.1307	0.0080	33.6097	36.1%
512	0.0553	19.1481	0.0163	65.1339	29.4%
1024	0.1782	23.6552	0.0601	70.1155	33.7%
1280	0.2852	23.0700	0.0829	79.3210	29.1%
2048	0.6455	26.0556	0.1967	85.4860	30.5%
4096	2.9182	23.0245	0.7721	87.0239	26.5%

Nvidia Geforce Gtx 1650 — beta = 2.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5759	0.0041	1.1654	49.4%
64	0.0086	2.0597	0.0044	4.0588	50.7%
128	0.0119	5.7281	0.0054	12.7045	45.1%
256	0.0220	12.1395	0.0074	36.3130	33.4%
512	0.0553	19.1481	0.0161	65.5857	29.2%
1024	0.1782	23.6552	0.0602	69.9665	33.8%
1280	0.2861	22.9965	0.0831	79.1835	29.0%
2048	0.6485	25.9335	0.1972	85.2987	30.4%
4096	2.9718	22.6093	0.7721	87.0239	26.0%

See also:

• beta.dsymv.js — WebGPU beta-sweep benchmark script
• beta.dsymv.c — CUDA / cuBLAS beta-sweep reference script

layout sweep

Column-major swaps the effective m/n and flips the transpose flag internally, changing which axis is contiguous and therefore how the matrix reads coalesce. wgblas-only: cuBLAS is column-major and has no layout argument, so there is no reference curve to compare against.

Nvidia Geforce Gtx 1650 — layout = column-major

n	compute ms	GB/s
32	0.0082	0.5759
64	0.0085	2.0791
128	0.0115	5.9276
256	0.0229	11.6730
512	0.0558	18.9616
1024	0.1781	23.6615
1280	0.2820	23.3278
2048	0.6386	26.3350
4096	3.1335	21.4429

Nvidia Geforce Gtx 1650 — layout = row-major

n	compute ms	GB/s
32	0.0082	0.5781
64	0.0086	2.0636
128	0.0114	5.9608
256	0.0220	12.1749
512	0.0553	19.1481
1024	0.1782	23.6552
1280	0.2849	23.0920
2048	0.6460	26.0363
4096	3.0294	22.1796

See also:

• layout.dsymv.js — WebGPU layout-sweep benchmark script