Module benchmarks/nvidia-geforce-gtx-1650/dsyr2

Benchmark results for dsyr2 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	0.0074	1.2148	0.0045	1.9858	61.2%
64	0.0074	4.6609	0.0045	7.5493	61.7%
128	0.0082	16.3750	0.0057	23.6169	69.3%
256	0.0123	43.1667	0.0084	62.7879	68.7%
512	0.0271	77.8736	0.0183	115.1441	67.6%
1024	0.0773	108.8888	0.0547	153.7048	70.8%
1280	0.1149	114.2984	0.0873	150.4434	76.0%
2048	0.2662	126.2154	0.2028	165.7243	76.2%
4096	1.0280	130.6556	0.8126	165.2933	79.0%

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

See also

• dsyr2.js — WebGPU benchmark script
• dsyr2.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	0.0088	1.0219	0.0049	1.8182	56.2%
64	0.0091	3.7880	0.0054	6.3432	59.7%
128	0.0107	12.5886	0.0061	21.8333	57.7%
256	0.0180	29.5209	0.0094	56.3810	52.4%
512	0.0410	51.5000	0.0184	114.4444	45.0%
1024	0.1208	69.6179	0.0555	151.5780	45.9%
1280	0.1802	72.8977	0.0895	146.7334	49.7%
2048	0.4301	78.1246	0.2067	162.5434	48.1%
4096	1.7221	77.9935	0.8520	157.6509	49.5%

Nvidia Geforce Gtx 1650 — stride = 5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0073	1.2335	0.0044	2.0217	61.0%
64	0.0078	4.3934	0.0047	7.3174	60.0%
128	0.0085	15.7594	0.0058	22.9699	68.6%
256	0.0140	37.9748	0.0089	59.6259	63.7%
512	0.0299	70.6538	0.0183	115.4466	61.2%
1024	0.0854	98.5058	0.0555	151.4906	65.0%
1280	0.1252	104.9489	0.0899	146.0808	71.8%
2048	0.3026	111.0466	0.2067	162.5434	68.3%
4096	1.2959	103.6466	0.8700	154.3777	67.1%

Nvidia Geforce Gtx 1650 — stride = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0075	1.1966	0.0056	1.5909	75.2%
64	0.0079	4.3226	0.0056	6.1257	70.6%
128	0.0115	11.6444	0.0062	21.6082	53.9%
256	0.0266	19.9231	0.0088	60.1670	33.1%
512	0.0757	27.8790	0.0193	109.3201	25.5%
1024	0.2662	31.6000	0.0578	145.4964	21.7%
1280	0.4043	32.4990	0.0981	133.9729	24.3%
2048	1.0041	33.4654	0.2263	148.4992	22.5%
4096	5.3321	25.1901	1.0361	129.6366	19.4%

Nvidia Geforce Gtx 1650 — stride = 33

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0074	1.2095	0.0052	1.7231	70.2%
64	0.0076	4.5232	0.0058	5.9227	76.4%
128	0.0089	15.0791	0.0062	21.7202	69.4%
256	0.0155	34.2126	0.0085	62.1989	55.0%
512	0.0466	45.2903	0.0187	112.8767	40.1%
1024	0.1802	46.6818	0.0573	146.8782	31.8%
1280	0.2799	46.9372	0.0983	133.6458	35.1%
2048	0.7148	47.0080	0.2279	147.4359	31.9%
4096	4.8842	27.5002	1.0407	129.0646	21.3%

Nvidia Geforce Gtx 1650 — stride = 255

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0075	1.1991	0.0052	1.7391	69.0%
64	0.0076	4.4854	0.0051	6.7210	66.7%
128	0.0091	14.7346	0.0061	22.0052	67.0%
256	0.0156	34.0370	0.0082	64.7500	52.6%
512	0.0471	44.7826	0.0186	113.3620	39.5%
1024	0.1839	45.7437	0.0570	147.7034	31.0%
1280	0.2847	46.1511	0.0999	131.5686	35.1%
2048	0.7434	45.2011	0.2292	146.5818	30.8%
4096	8.8679	15.1464	1.1033	121.7442	12.4%

Nvidia Geforce Gtx 1650 — stride = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0074	1.2121	0.0054	1.6568	73.2%
64	0.0082	4.1875	0.0058	5.9227	70.7%
128	0.0115	11.6283	0.0061	21.8333	53.3%
256	0.0265	20.0314	0.0092	57.6557	34.7%
512	0.0754	27.9618	0.0196	107.4491	26.0%
1024	0.2662	31.6000	0.0572	147.0836	21.5%
1280	0.4055	32.3990	0.0977	134.4776	24.1%
2048	0.9885	33.9958	0.2270	148.0282	23.0%
4096	7.6562	17.5434	1.1084	121.1836	14.5%

See also:

• stride.dsyr2.js — WebGPU stride-sweep benchmark script
• stride.dsyr2.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	0.0072	1.2362	0.0054	1.6568	74.6%
64	0.0073	4.6710	0.0056	6.0909	76.7%
128	0.0083	16.0921	0.0062	21.6641	74.3%
256	0.0131	40.6275	0.0094	56.5734	71.8%
512	0.0290	72.7594	0.0181	116.4664	62.5%
1024	0.0799	105.3333	0.0553	152.1481	69.2%
1280	0.1166	112.6829	0.0886	148.3237	76.0%
2048	0.2728	123.1658	0.2068	162.4554	75.8%
4096	1.0177	131.9847	0.8192	163.9600	80.5%

Nvidia Geforce Gtx 1650 — uplo = upper

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0074	1.2043	0.0055	1.6185	74.4%
64	0.0075	4.5520	0.0046	7.3931	61.6%
128	0.0082	16.3750	0.0059	22.5984	72.5%
256	0.0123	43.1667	0.0083	63.6315	67.8%
512	0.0270	78.1043	0.0185	113.8515	68.6%
1024	0.0767	109.7525	0.0555	151.5780	72.4%
1280	0.1150	114.2825	0.0912	144.0309	79.3%
2048	0.2664	126.1244	0.2069	162.4177	77.7%
4096	1.0268	130.8144	0.8191	163.9792	79.8%

See also:

• uplo.dsyr2.js — WebGPU uplo-sweep benchmark script
• uplo.dsyr2.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	0.0071	1.2556	0.0045	2.0000	62.8%
64	0.0074	4.6207	0.0046	7.3931	62.5%
128	0.0083	16.1853	0.0056	23.8182	68.0%
256	0.0135	39.2331	0.0082	64.7500	60.6%
512	0.0287	73.4485	0.0182	115.8524	63.4%
1024	0.0799	105.3333	0.0553	152.2363	69.2%
1280	0.1167	112.5439	0.0873	150.4158	74.8%
2048	0.2736	122.8201	0.2028	165.6851	74.1%
4096	1.0162	132.1696	0.8129	165.2250	80.0%

Nvidia Geforce Gtx 1650 — pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0071	1.2698	0.0045	1.9718	64.4%
64	0.0075	4.5520	0.0046	7.4444	61.1%
128	0.0084	16.0000	0.0055	24.3721	65.6%
256	0.0134	39.4667	0.0082	64.7500	61.0%
512	0.0295	71.4193	0.0184	114.4444	62.4%
1024	0.0810	103.9178	0.0692	121.4935	85.5%
1280	0.1188	110.6034	0.1249	105.1639	105.2%
2048	0.2817	119.2970	0.2704	124.2663	96.0%

Nvidia Geforce Gtx 1650 — pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0072	1.2500	0.0044	2.0588	60.7%
64	0.0073	4.7225	0.0046	7.4965	63.0%
128	0.0083	16.0921	0.0055	24.5865	65.5%
256	0.0135	39.3729	0.0081	65.2598	60.3%
512	0.0289	72.9204	0.0184	114.4444	63.7%
1024	0.0801	105.0597	0.0629	133.7294	78.6%
1280	0.1200	109.5119	0.1161	113.1799	96.8%
2048	0.2757	121.8793	0.2924	114.9422	106.0%

Nvidia Geforce Gtx 1650 — pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0071	1.2556	0.0044	2.0144	62.3%
64	0.0075	4.5910	0.0048	7.2189	63.6%
128	0.0084	16.0000	0.0055	24.5146	65.3%
256	0.0133	39.9903	0.0082	64.7500	61.8%
512	0.0291	72.5991	0.0177	118.8819	61.1%
1024	0.0799	105.3333	0.0589	142.7706	73.8%
1280	0.1188	110.6034	0.1114	117.9770	93.8%
2048	0.2802	119.9237	0.2593	129.5874	92.5%

Nvidia Geforce Gtx 1650 — pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0087	1.0256	0.0045	2.0000	51.3%
64	0.0089	3.8631	0.0046	7.4444	51.9%
128	0.0103	13.0389	0.0060	22.4171	58.2%
256	0.0163	32.6299	0.0088	59.9494	54.4%
512	0.0360	58.5175	0.0174	121.3996	48.2%
1024	0.1004	83.8367	0.0564	149.1276	56.2%
1280	0.1472	89.2425	0.0855	153.6814	58.1%
2048	0.3453	97.3227	0.2189	153.5251	63.4%

Nvidia Geforce Gtx 1650 — pad = 48

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0087	1.0351	0.0049	1.8301	56.6%
64	0.0091	3.7880	0.0045	7.5760	50.0%
128	0.0104	12.9383	0.0060	22.2387	58.2%
256	0.0162	32.7589	0.0084	62.7879	52.2%
512	0.0483	43.7135	0.0182	115.8524	37.7%
1024	0.1393	60.4118	0.0646	130.3158	46.4%
1280	0.2096	62.6857	0.1081	121.5214	51.6%
2048	0.4894	68.6572	0.2747	122.3479	56.1%

Nvidia Geforce Gtx 1650 — pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0124	0.7254	0.0051	1.7666	41.1%
64	0.0128	2.6733	0.0048	7.1467	37.4%
128	0.0160	8.4008	0.0061	22.0632	38.1%
256	0.0243	21.8537	0.0085	62.5509	34.9%
512	0.0552	38.1924	0.0177	119.4203	32.0%
1024	0.1577	53.3506	0.0576	146.0217	36.5%
1280	0.2429	54.0815	0.0867	151.5821	35.7%
2048	0.5898	56.9738	0.2150	156.2899	36.5%

Nvidia Geforce Gtx 1650 — pad = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0137	0.6542	0.0054	1.6716	39.1%
64	0.0143	2.4036	0.0057	6.0565	39.7%
128	0.0169	7.9169	0.0061	21.8333	36.3%
256	0.0266	19.9231	0.0083	64.0000	31.1%
512	0.0594	35.5364	0.0184	114.4444	31.1%
1024	0.1734	48.5122	0.0600	140.2197	34.6%
1280	0.2601	50.5118	0.0920	142.8283	35.4%
2048	0.6498	51.7118	0.2334	143.9594	35.9%

See also:

• lda.dsyr2.js — WebGPU lda-sweep benchmark script
• lda.dsyr2.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	0.0074	1.2121	0.0045	2.0000	60.6%
64	0.0079	4.3577	0.0046	7.4965	58.1%
128	0.0083	16.1231	0.0057	23.6836	68.1%
256	0.0136	39.1405	0.0082	64.7500	60.4%
512	0.0291	72.4794	0.0184	114.4444	63.3%
1024	0.0799	105.3333	0.0545	154.3364	68.2%
1280	0.1167	112.5439	0.0875	150.1957	74.9%
2048	0.2754	122.0351	0.2028	165.6851	73.7%
4096	1.0172	132.0511	0.8126	165.2933	79.9%

Nvidia Geforce Gtx 1650 — alpha = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0073	1.2308	0.0016	5.4369	22.6%
64	0.0074	4.6207	0.0016	21.6566	21.3%
128	0.0084	15.9392	0.0016	85.5510	18.6%
256	0.0148	35.9566	0.0016	328.2376	11.0%
512	0.0291	72.5592	0.0016	1318.4001	5.5%
1024	0.0799	105.3333	0.0020	4240.5166	2.5%
1280	0.1165	112.7602	0.0015	8643.3682	1.3%
2048	0.2730	123.1081	0.0016	21002.2402	0.6%
4096	1.0179	131.9598	0.0018	74953.1406	0.2%

Nvidia Geforce Gtx 1650 — alpha = 1e-38

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0075	1.2017	0.0047	1.9048	63.1%
64	0.0077	4.4389	0.0051	6.7848	65.4%
128	0.0084	15.8788	0.0056	23.8860	66.5%
256	0.0136	39.0943	0.0082	64.7500	60.4%
512	0.0292	72.3600	0.0182	116.1586	62.3%
1024	0.0799	105.3333	0.0548	153.5253	68.6%
1280	0.1165	112.7912	0.0873	150.5537	74.9%
2048	0.2716	123.7391	0.2028	165.7112	74.7%
4096	1.0159	132.2091	0.8123	165.3617	80.0%

Nvidia Geforce Gtx 1650 — alpha = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0071	1.2698	0.0061	1.4583	87.1%
64	0.0081	4.2288	0.0060	5.6720	74.6%
128	0.0084	16.0000	0.0068	19.7736	80.9%
256	0.0136	39.0024	0.0102	51.8000	75.3%
512	0.0292	72.3600	0.0224	94.3062	76.7%
1024	0.0798	105.4178	0.0675	124.6619	84.6%
1280	0.1167	112.5439	0.1018	129.0461	87.2%
2048	0.2724	123.3684	0.2494	134.7507	91.6%
4096	1.0239	131.1823	0.9710	138.3332	94.8%

Nvidia Geforce Gtx 1650 — alpha = 2.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0075	1.1991	0.0058	1.5427	77.7%
64	0.0080	4.3139	0.0059	5.8420	73.8%
128	0.0083	16.0921	0.0067	19.9619	80.6%
256	0.0131	40.4786	0.0102	51.8000	78.1%
512	0.0290	72.6792	0.0182	116.0563	62.6%
1024	0.0797	105.5448	0.0549	153.2568	68.9%
1280	0.1167	112.5747	0.0877	149.8394	75.1%
2048	0.2724	123.3684	0.2028	165.7374	74.4%
4096	1.0164	132.1509	0.8130	165.2185	80.0%

See also:

• alpha.dsyr2.js — WebGPU alpha-sweep benchmark script
• alpha.dsyr2.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	0.0076	1.1814
64	0.0075	4.5520
128	0.0082	16.3750
256	0.0123	43.2229
512	0.0271	77.9196
1024	0.0769	109.4099
1280	0.1163	112.9464
2048	0.2676	125.5889
4096	1.0237	131.2029

Nvidia Geforce Gtx 1650 — layout = row-major

n	compute ms	GB/s
32	0.0095	0.9459
64	0.0079	4.3666
128	0.0083	16.1231
256	0.0139	38.0620
512	0.0292	72.3600
1024	0.0799	105.3333
1280	0.1166	112.6365
2048	0.2729	123.1514
4096	1.0163	132.1675

See also:

• layout.dsyr2.js — WebGPU layout-sweep benchmark script