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

Benchmark results for dsyr 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	0.0076	1.1429	0.0041	2.1250	53.8%
64	0.0082	4.1250	0.0039	8.6914	47.5%
128	0.0083	16.0000	0.0041	32.5000	49.2%
256	0.0102	51.6000	0.0058	91.7333	56.3%
512	0.0200	105.4359	0.0134	156.6476	67.3%
1024	0.0631	133.2941	0.0530	158.7045	84.0%
1280	0.1024	128.2000	0.0882	148.7994	86.2%
2048	0.2188	153.5289	0.2010	167.0805	91.9%
4096	0.8848	151.7586	0.8187	164.0289	92.5%

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

See also

• dsyr.js — WebGPU benchmark script
• dsyr.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	1.0625	0.0037	2.3652	44.9%
64	0.0083	4.0772	0.0039	8.7273	46.7%
128	0.0095	13.9597	0.0042	31.7557	44.0%
256	0.0139	38.0900	0.0060	88.2995	43.1%
512	0.0281	74.8061	0.0133	158.1539	47.3%
1024	0.0737	114.0000	0.0531	158.2265	72.0%
1280	0.1085	121.0147	0.0867	151.4360	79.9%
2048	0.2519	133.3333	0.2011	167.0008	79.8%
4096	0.9359	143.4748	0.8266	162.4572	88.3%

Nvidia Geforce Gtx 1650 — stride = 5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0068	1.2891	0.0037	2.3550	54.7%
64	0.0071	4.7568	0.0038	8.8000	54.1%
128	0.0081	16.5079	0.0041	32.5000	50.8%
256	0.0102	51.6000	0.0058	90.7253	56.9%
512	0.0220	95.4891	0.0135	155.7207	61.3%
1024	0.0635	132.3871	0.0530	158.5130	83.5%
1280	0.0953	137.7338	0.0874	150.2711	91.7%
2048	0.2213	151.7969	0.2010	167.1071	90.8%
4096	0.8479	158.3768	0.8296	161.8744	97.8%

Nvidia Geforce Gtx 1650 — stride = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0068	1.2830	0.0037	2.3248	55.2%
64	0.0073	4.6316	0.0038	8.9872	51.5%
128	0.0092	14.4194	0.0041	32.2481	44.7%
256	0.0175	30.2141	0.0061	86.4503	34.9%
512	0.0436	48.3232	0.0140	150.2101	32.2%
1024	0.1495	56.2192	0.0532	157.9411	35.6%
1280	0.2670	49.1718	0.0915	143.4657	34.3%
2048	0.6805	49.3581	0.2042	164.4754	30.0%
4096	2.5703	52.2451	0.9254	145.1046	36.0%

Nvidia Geforce Gtx 1650 — stride = 33

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0067	1.2952	0.0037	2.3248	55.7%
64	0.0068	5.0047	0.0036	9.3451	53.6%
128	0.0080	16.5737	0.0042	32.0000	51.8%
256	0.0110	48.1399	0.0061	86.0000	56.0%
512	0.0268	78.4639	0.0140	150.2101	52.2%
1024	0.1004	83.7551	0.0532	157.8461	53.1%
1280	0.1556	84.3421	0.0922	142.4444	59.2%
2048	0.3975	84.4953	0.2053	163.5655	51.7%
4096	1.6056	83.6327	0.9400	142.8497	58.5%

Nvidia Geforce Gtx 1650 — stride = 255

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0067	1.3014	0.0037	2.3550	55.3%
64	0.0071	4.7568	0.0038	8.8739	53.6%
128	0.0080	16.5737	0.0041	32.5000	51.0%
256	0.0109	48.3514	0.0061	86.0000	56.2%
512	0.0267	78.7928	0.0137	153.3613	51.4%
1024	0.1024	82.0800	0.0532	157.9411	52.0%
1280	0.1587	82.7013	0.0921	142.5434	58.0%
2048	0.4018	83.5968	0.2048	164.0256	51.0%
4096	1.9178	70.0203	0.9522	141.0207	49.7%

Nvidia Geforce Gtx 1650 — stride = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0067	1.2952	0.0039	2.2295	58.1%
64	0.0072	4.7143	0.0040	8.4143	56.0%
128	0.0093	14.2466	0.0042	31.7557	44.9%
256	0.0172	30.6914	0.0059	89.7391	34.2%
512	0.0432	48.6807	0.0141	149.5273	32.6%
1024	0.1475	56.9876	0.0532	158.0837	36.0%
1280	0.2238	58.6602	0.0922	142.4444	41.2%
2048	0.5475	61.3425	0.2045	164.2053	37.4%
4096	2.1299	63.0480	0.9555	140.5319	44.9%

See also:

• stride.dsyr.js — WebGPU stride-sweep benchmark script
• stride.dsyr.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	0.0066	1.3204	0.0036	2.3860	55.3%
64	0.0069	4.9231	0.0037	9.1034	54.1%
128	0.0076	17.4423	0.0041	32.5000	53.7%
256	0.0101	52.1706	0.0058	91.7333	56.9%
512	0.0210	100.1400	0.0130	162.2491	61.7%
1024	0.0614	136.8000	0.0531	158.4174	86.4%
1280	0.0928	141.4621	0.0870	150.8790	93.8%
2048	0.2172	154.6144	0.2006	167.4003	92.4%
4096	0.8274	162.2908	0.8187	164.0193	98.9%

Nvidia Geforce Gtx 1650 — uplo = upper

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0066	1.3236	0.0040	2.1587	61.3%
64	0.0067	5.0286	0.0040	8.5161	59.0%
128	0.0076	17.4790	0.0041	32.2481	54.2%
256	0.0097	54.6755	0.0058	91.7333	59.6%
512	0.0190	110.5748	0.0135	156.4614	70.7%
1024	0.0631	133.1927	0.0532	157.8461	84.4%
1280	0.1004	130.8163	0.0871	150.6574	86.8%
2048	0.2239	150.0286	0.2007	167.3469	89.7%
4096	0.8519	157.6213	0.8172	164.3308	95.9%

See also:

• uplo.dsyr.js — WebGPU uplo-sweep benchmark script
• uplo.dsyr.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	0.0066	1.3204	0.0037	2.3550	56.1%
64	0.0068	4.9577	0.0037	9.0644	54.7%
128	0.0077	17.2614	0.0041	32.5000	53.1%
256	0.0101	52.2532	0.0058	91.7333	57.0%
512	0.0211	99.7604	0.0129	162.8515	61.3%
1024	0.0614	136.8000	0.0529	158.8005	86.1%
1280	0.0932	140.8308	0.0867	151.4360	93.0%
2048	0.2173	154.5347	0.2011	167.0008	92.5%
4096	0.8311	161.5722	0.8211	163.5367	98.8%

Nvidia Geforce Gtx 1650 — pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0066	1.3204	0.0038	2.2857	57.8%
64	0.0067	5.0526	0.0041	8.2500	61.2%
128	0.0078	17.0143	0.0044	29.9281	56.9%
256	0.0102	51.6000	0.0063	83.3939	61.9%
512	0.0220	95.9067	0.0184	114.3215	83.9%
1024	0.0655	128.2500	0.0672	125.0148	102.6%
1280	0.1066	123.1582	0.1262	104.0030	118.4%
2048	0.2334	143.9287	0.2601	129.1339	111.5%

Nvidia Geforce Gtx 1650 — pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0068	1.2800	0.0037	2.3652	54.1%
64	0.0068	5.0047	0.0041	8.2500	60.7%
128	0.0076	17.4790	0.0041	32.3735	54.0%
256	0.0102	51.6000	0.0057	93.0254	55.5%
512	0.0216	97.6142	0.0184	114.2222	85.5%
1024	0.0655	128.2500	0.0595	141.3269	90.7%
1280	0.1088	120.6056	0.1176	111.5996	108.1%
2048	0.2377	141.2746	0.3012	111.4994	126.7%

Nvidia Geforce Gtx 1650 — pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0066	1.3268	0.0038	2.2954	57.8%
64	0.0067	5.0526	0.0037	9.1034	55.5%
128	0.0078	17.0842	0.0041	32.3735	52.8%
256	0.0102	51.6000	0.0059	90.2295	57.2%
512	0.0216	97.6142	0.0143	146.8571	66.5%
1024	0.0636	132.1540	0.0574	146.4897	90.2%
1280	0.1020	128.7027	0.1102	119.0827	108.1%
2048	0.2458	136.6400	0.2540	132.2581	103.3%

Nvidia Geforce Gtx 1650 — pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	1.0625	0.0037	2.3248	45.7%
64	0.0082	4.1089	0.0037	9.1429	44.9%
128	0.0095	13.9832	0.0041	32.5000	43.0%
256	0.0125	42.2302	0.0057	92.7640	45.5%
512	0.0259	81.2749	0.0143	146.8571	55.3%
1024	0.0835	100.6538	0.0541	155.2799	64.8%
1280	0.1226	107.0843	0.0842	155.9255	68.7%
2048	0.2805	119.7353	0.2149	156.2719	76.6%

Nvidia Geforce Gtx 1650 — pad = 48

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0107	0.8144	0.0041	2.1333	38.2%
64	0.0111	3.0476	0.0038	8.8739	34.3%
128	0.0124	10.7216	0.0041	32.5000	33.0%
256	0.0174	30.4369	0.0057	93.0254	32.7%
512	0.0350	60.0840	0.0163	129.5118	46.4%
1024	0.0932	90.1668	0.0655	128.2500	70.3%
1280	0.1326	98.9723	0.1082	121.3010	81.6%
2048	0.3181	105.5829	0.2681	125.2655	84.3%

Nvidia Geforce Gtx 1650 — pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0121	0.7167	0.0037	2.3248	30.8%
64	0.0123	2.7500	0.0037	9.1826	29.9%
128	0.0137	9.7196	0.0042	31.7557	30.6%
256	0.0185	28.6170	0.0056	93.8182	30.5%
512	0.0383	55.0100	0.0142	148.5147	37.0%
1024	0.1021	82.3373	0.0573	146.5714	56.2%
1280	0.1450	90.5607	0.0847	155.0123	58.4%
2048	0.3354	100.1288	0.2086	161.0310	62.2%

Nvidia Geforce Gtx 1650 — pad = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0125	0.6983	0.0037	2.3652	29.5%
64	0.0133	2.5415	0.0037	9.1429	27.8%
128	0.0143	9.2857	0.0041	32.5000	28.6%
256	0.0202	26.1473	0.0057	93.2881	28.0%
512	0.0408	51.5409	0.0143	146.8571	35.1%
1024	0.1095	76.7776	0.0573	146.6533	52.4%
1280	0.1591	82.5184	0.0878	149.5589	55.2%
2048	0.3707	90.6077	0.2331	144.0966	62.9%

See also:

• lda.dsyr.js — WebGPU lda-sweep benchmark script
• lda.dsyr.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	0.0081	1.0794	0.0040	2.1847	49.4%
64	0.0082	4.1250	0.0038	8.9872	45.9%
128	0.0089	14.8837	0.0041	32.2481	46.2%
256	0.0124	42.4473	0.0061	86.4503	49.1%
512	0.0226	93.2558	0.0154	136.6397	68.2%
1024	0.0614	136.8356	0.0530	158.6566	86.2%
1280	0.0939	139.8704	0.0876	149.8320	93.4%
2048	0.2175	154.4324	0.2020	166.2733	92.9%
4096	0.8294	161.8963	0.8181	164.1444	98.6%

Nvidia Geforce Gtx 1650 — alpha = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0096	0.9097	0.0016	5.3861	16.9%
64	0.0097	3.4909	0.0013	25.1429	13.9%
128	0.0106	12.5680	0.0014	94.5455	13.3%
256	0.0140	37.8716	0.0013	393.1429	9.6%
512	0.0265	79.4589	0.0014	1495.2728	5.3%
1024	0.0657	127.9377	0.0014	5902.3813	2.2%
1280	0.1004	130.8163	0.0016	8372.2451	1.6%
2048	0.2281	147.2606	0.0014	23324.4434	0.6%
4096	0.8476	158.4187	0.0015	89284.0781	0.2%

Nvidia Geforce Gtx 1650 — alpha = 1e-38

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	1.0625	0.0038	2.2667	46.9%
64	0.0080	4.1988	0.0039	8.6914	48.3%
128	0.0090	14.8571	0.0042	31.6350	47.0%
256	0.0126	42.0688	0.0061	86.6772	48.5%
512	0.0226	92.9922	0.0140	149.8679	62.0%
1024	0.0614	136.8000	0.0529	158.8965	86.1%
1280	0.0940	139.6324	0.0871	150.6850	92.7%
2048	0.2175	154.4097	0.2013	166.8150	92.6%
4096	0.8294	161.8963	0.8183	164.1091	98.7%

Nvidia Geforce Gtx 1650 — alpha = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	1.0625	0.0038	2.2667	46.9%
64	0.0084	4.0152	0.0038	8.9872	44.7%
128	0.0102	13.0000	0.0042	31.5152	41.2%
256	0.0143	36.8571	0.0060	88.7742	41.5%
512	0.0249	84.6744	0.0158	132.9131	63.7%
1024	0.0637	131.9879	0.0531	158.3219	83.4%
1280	0.0981	133.8030	0.0866	151.5758	88.3%
2048	0.2223	151.0868	0.2018	166.4446	90.8%
4096	0.8330	161.2090	0.8204	163.6706	98.5%

Nvidia Geforce Gtx 1650 — alpha = 2.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	1.0625	0.0143	0.6071	175.0%
64	0.0082	4.1250	0.0048	7.1111	58.0%
128	0.0090	14.7257	0.0054	24.6884	59.6%
256	0.0126	41.9619	0.0057	92.2458	45.5%
512	0.0227	92.6648	0.0129	163.2556	56.8%
1024	0.0612	137.3365	0.0532	157.8461	87.0%
1280	0.0939	139.8466	0.0872	150.5744	92.9%
2048	0.2176	154.3756	0.2018	166.4183	92.8%
4096	0.8297	161.8401	0.8185	164.0546	98.7%

See also:

• alpha.dsyr.js — WebGPU alpha-sweep benchmark script
• alpha.dsyr.c — CUDA / cuBLAS alpha-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	1.0625
64	0.0071	4.7568
128	0.0077	17.2973
256	0.0110	48.1399
512	0.0211	99.9878
1024	0.0635	132.3871
1280	0.1018	128.9252
2048	0.2239	150.0286
4096	0.8516	157.6924

Nvidia Geforce Gtx 1650 — layout = row-major

n	compute ms	GB/s
32	0.0066	1.3204
64	0.0072	4.6726
128	0.0091	14.5965
256	0.0126	42.0153
512	0.0211	99.6848
1024	0.0614	136.8000
1280	0.0940	139.7275
2048	0.2190	153.3382
4096	0.8315	161.4975

See also:

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