Module benchmarks/nvidia-geforce-gtx-1650/ssyr2

Benchmark results for ssyr2 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	0.0061	0.7292	0.0037	1.2174	59.9%
64	0.0061	2.7917	0.0041	4.1875	66.7%
128	0.0064	10.5327	0.0039	17.2510	61.1%
256	0.0072	36.5916	0.0044	60.9412	60.0%
512	0.0097	108.4211	0.0072	146.4889	74.0%
1024	0.0559	75.2683	0.0332	126.5217	59.5%
1280	0.0930	70.6522	0.0695	94.5120	74.8%
2048	0.2294	73.2500	0.1190	141.2200	51.9%
4096	0.8366	80.2757	0.4790	140.2023	57.3%

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

See also

• ssyr2.js — WebGPU benchmark script
• ssyr2.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	0.0076	0.5932	0.0036	1.2389	47.9%
64	0.0080	2.1397	0.0037	4.5812	46.7%
128	0.0082	8.1875	0.0038	17.7627	46.1%
256	0.0097	27.2632	0.0042	63.7538	42.8%
512	0.0143	73.5714	0.0068	154.7418	47.5%
1024	0.0389	108.1053	0.0440	95.6391	113.0%
1280	0.0578	113.5712	0.0695	94.5773	120.1%
2048	0.1597	105.1795	0.1272	132.1394	79.6%
4096	0.7999	83.9576	0.4895	137.2050	61.2%

Nvidia Geforce Gtx 1650 — stride = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0075	0.5957	0.0038	1.1864	50.2%
64	0.0081	2.1228	0.0038	4.5424	46.7%
128	0.0094	7.1658	0.0039	17.3223	41.4%
256	0.0164	16.1402	0.0047	56.7671	28.4%
512	0.0395	26.6775	0.0082	128.7500	20.7%
1024	0.1495	28.1370	0.0353	119.0725	23.6%
1280	0.2298	28.5905	0.0710	92.4893	30.9%
2048	0.5595	30.0324	0.1314	127.8442	23.5%
4096	1.8779	35.7631	0.6141	109.3636	32.7%

Nvidia Geforce Gtx 1650 — stride = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7292	0.0039	1.1475	63.5%
64	0.0066	2.6146	0.0039	4.4298	59.0%
128	0.0079	8.4858	0.0039	17.3223	49.0%
256	0.0142	18.6247	0.0048	55.6242	33.5%
512	0.0313	33.7359	0.0081	129.7638	26.0%
1024	0.1250	33.6592	0.0350	120.3259	28.0%
1280	0.1928	34.0799	0.0696	94.3382	36.1%
2048	0.4760	35.2955	0.1352	124.3030	28.4%
4096	1.8876	35.5791	0.6349	105.7806	33.6%

See also:

• stride.ssyr2.js — WebGPU stride-sweep benchmark script
• stride.ssyr2.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	0.0061	0.7292	0.0036	1.2500	58.3%
64	0.0061	2.7917	0.0041	4.2205	66.1%
128	0.0064	10.5327	0.0038	17.4667	60.3%
256	0.0076	35.1186	0.0041	64.2481	54.7%
512	0.0096	109.8667	0.0067	156.9524	70.0%
1024	0.0335	125.6149	0.0326	129.1317	97.3%
1280	0.0512	128.3000	0.0677	97.0132	132.3%
2048	0.1148	146.3979	0.1260	133.2968	109.8%
4096	0.4342	154.6792	0.4664	143.9874	107.4%

Nvidia Geforce Gtx 1650 — uplo = upper

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7292	0.0035	1.2785	57.0%
64	0.0062	2.7844	0.0036	4.7857	58.2%
128	0.0064	10.5327	0.0038	17.6134	59.8%
256	0.0074	35.9566	0.0041	64.7500	55.5%
512	0.0099	106.1514	0.0070	151.5402	70.0%
1024	0.0524	80.2785	0.0331	127.0106	63.2%
1280	0.0893	73.5244	0.0676	97.1970	75.6%
2048	0.2068	81.2277	0.1262	133.0940	61.0%
4096	0.7737	86.7980	0.4590	146.3010	59.3%

See also:

• uplo.ssyr2.js — WebGPU uplo-sweep benchmark script
• uplo.ssyr2.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	0.0077	0.5797	0.0036	1.2500	46.4%
64	0.0077	2.2333	0.0037	4.5812	48.7%
128	0.0080	8.3673	0.0039	17.3223	48.3%
256	0.0091	29.0298	0.0043	61.8507	46.9%
512	0.0120	87.7763	0.0070	151.5402	57.9%
1024	0.0367	114.6085	0.0328	128.3750	89.3%
1280	0.0544	120.6820	0.0689	95.3018	126.6%
2048	0.1206	139.2906	0.1189	141.2770	98.6%
4096	0.4444	151.1152	0.4813	139.5451	108.3%

Nvidia Geforce Gtx 1650 — pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0075	0.6009	0.0038	1.1864	50.6%
64	0.0078	2.2058	0.0038	4.4667	49.4%
128	0.0082	8.1875	0.0038	17.4667	46.9%
256	0.0093	28.4811	0.0054	49.0414	58.1%
512	0.0123	85.8333	0.0073	144.5614	59.4%
1024	0.0614	68.4667	0.0409	102.7401	66.6%
1280	0.0965	68.0411	0.0778	84.4253	80.6%
2048	0.2320	72.4165	0.1553	108.1920	66.9%
4096	0.8857	75.8279	0.7662	87.6517	86.5%

Nvidia Geforce Gtx 1650 — pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0064	0.6983	0.0043	1.0448	66.8%
64	0.0065	2.6210	0.0041	4.1875	62.6%
128	0.0071	9.5057	0.0040	16.9032	56.2%
256	0.0082	32.5020	0.0045	58.9893	55.1%
512	0.0103	102.3602	0.0076	138.7789	73.8%
1024	0.0568	74.0806	0.0445	94.6067	78.3%
1280	0.0927	70.8717	0.0758	86.6160	81.8%
2048	0.2156	77.9188	0.1817	92.4720	84.3%
4096	0.8132	82.5865	0.7572	88.6944	93.1%

Nvidia Geforce Gtx 1650 — pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0065	0.6880	0.0037	1.2069	57.0%
64	0.0065	2.6210	0.0036	4.7857	54.8%
128	0.0073	9.2132	0.0036	18.3860	50.1%
256	0.0082	32.3750	0.0042	62.7879	51.6%
512	0.0102	103.0000	0.0068	155.4717	66.2%
1024	0.0573	73.3571	0.0451	93.3636	78.6%
1280	0.0911	72.1040	0.0708	92.8239	77.7%
2048	0.2199	76.4051	0.1494	112.4919	67.9%
4096	0.8131	82.5995	0.7872	85.3091	96.8%

Nvidia Geforce Gtx 1650 — pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7292	0.0035	1.2670	57.6%
64	0.0063	2.7347	0.0036	4.7857	57.1%
128	0.0065	10.3251	0.0037	18.2261	56.7%
256	0.0079	33.6227	0.0041	64.7500	51.9%
512	0.0097	108.2430	0.0067	157.7034	68.6%
1024	0.0346	121.4374	0.0395	106.6148	113.9%
1280	0.0512	128.2599	0.0604	108.6713	118.0%
2048	0.1154	145.6062	0.1433	117.2262	124.2%
4096	0.4483	149.7993	0.6189	108.5154	138.0%

Nvidia Geforce Gtx 1650 — pad = 48

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7292	0.0036	1.2281	59.4%
64	0.0061	2.7917	0.0037	4.6207	60.4%
128	0.0066	10.2244	0.0037	18.2261	56.1%
256	0.0078	34.1773	0.0041	64.2481	53.2%
512	0.0097	108.7789	0.0070	150.5023	72.3%
1024	0.0585	71.9321	0.0462	91.0991	79.0%
1280	0.0880	74.6608	0.0695	94.5337	79.0%
2048	0.2191	76.6953	0.1677	100.1633	76.6%
4096	0.8149	82.4130	0.8247	81.4313	101.2%

Nvidia Geforce Gtx 1650 — pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7292	0.0037	1.2174	59.9%
64	0.0061	2.7917	0.0038	4.5424	61.5%
128	0.0065	10.2745	0.0041	16.3750	62.7%
256	0.0077	34.4615	0.0042	62.5509	55.1%
512	0.0096	109.5017	0.0068	156.2085	70.1%
1024	0.0348	120.7126	0.0431	97.7005	123.6%
1280	0.0506	129.8829	0.0593	110.7825	117.2%
2048	0.1191	141.0493	0.1560	107.7038	131.0%
4096	0.4440	151.2677	0.6017	111.6145	135.5%

Nvidia Geforce Gtx 1650 — pad = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7292	0.0042	1.0566	69.0%
64	0.0063	2.7347	0.0041	4.1712	65.6%
128	0.0065	10.2745	0.0043	15.6418	65.7%
256	0.0075	35.2681	0.0044	59.6259	59.1%
512	0.0100	105.3035	0.0070	149.8182	70.3%
1024	0.0346	121.4374	0.0428	98.3952	123.4%
1280	0.0512	128.3000	0.0630	104.2560	123.1%
2048	0.1188	141.4483	0.1909	87.9933	160.7%
4096	0.4873	137.8268	0.6371	105.4061	130.8%

See also:

• lda.ssyr2.js — WebGPU lda-sweep benchmark script
• lda.ssyr2.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	0.6061	0.0043	1.0448	58.0%
64	0.0077	2.2149	0.0040	4.2709	51.9%
128	0.0079	8.4858	0.0041	16.3750	51.8%
256	0.0093	28.5301	0.0047	56.7671	50.3%
512	0.0120	88.2463	0.0076	137.9080	64.0%
1024	0.0368	114.3096	0.0328	128.0623	89.3%
1280	0.0544	120.8596	0.0680	96.5796	125.1%
2048	0.1204	139.5869	0.1267	132.6569	105.2%
4096	0.4451	150.8871	0.4814	139.5172	108.1%

Nvidia Geforce Gtx 1650 — alpha = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7292	0.0014	3.1111	23.4%
64	0.0061	2.7917	0.0015	11.2842	24.7%
128	0.0064	10.5327	0.0016	41.9200	25.1%
256	0.0075	35.1932	0.0015	178.2365	19.7%
512	0.0096	109.8667	0.0014	749.0909	14.7%
1024	0.0334	126.0364	0.0018	2390.1091	5.3%
1280	0.0512	128.3000	0.0015	4276.6665	3.0%
2048	0.1149	146.2144	0.0014	11799.0098	1.2%
4096	0.4342	154.6849	0.0017	38508.0352	0.4%

Nvidia Geforce Gtx 1650 — alpha = 1e-38

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0075	0.5983	0.0044	1.0182	58.8%
64	0.0077	2.2333	0.0040	4.2880	52.1%
128	0.0081	8.2358	0.0041	16.3750	50.3%
256	0.0093	28.5301	0.0049	53.6440	53.2%
512	0.0118	89.0811	0.0076	139.3658	63.9%
1024	0.0369	114.1111	0.0328	128.3750	88.9%
1280	0.0545	120.4695	0.0678	96.9445	124.3%
2048	0.1204	139.4942	0.1259	133.4832	104.5%
4096	0.5176	129.7449	0.4813	139.5404	93.0%

Nvidia Geforce Gtx 1650 — alpha = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7292	0.0051	0.8723	83.6%
64	0.0061	2.7917	0.0045	3.8286	72.9%
128	0.0064	10.5592	0.0045	14.9714	70.5%
256	0.0077	34.5333	0.0052	50.8466	67.9%
512	0.0096	109.8667	0.0083	127.0135	86.5%
1024	0.0334	126.0969	0.0328	128.3750	98.2%
1280	0.0512	128.3000	0.0678	96.8302	132.5%
2048	0.1149	146.2552	0.1270	132.3226	110.5%
4096	0.4404	152.4766	0.4813	139.5404	109.3%

Nvidia Geforce Gtx 1650 — alpha = 2.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0075	0.5945	0.0044	1.0256	58.0%
64	0.0077	2.2380	0.0041	4.1875	53.4%
128	0.0079	8.4516	0.0041	16.3750	51.6%
256	0.0092	28.8278	0.0045	58.3662	49.4%
512	0.0120	87.8933	0.0072	145.5188	60.4%
1024	0.0367	114.7586	0.0327	128.5005	89.3%
1280	0.0540	121.5394	0.0678	96.8302	125.5%
2048	0.1203	139.7169	0.1269	132.4227	105.5%
4096	0.4447	151.0228	0.4813	139.5404	108.2%

See also:

• alpha.ssyr2.js — WebGPU alpha-sweep benchmark script
• alpha.ssyr2.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	0.5858
64	0.0079	2.1789
128	0.0079	8.4516
256	0.0085	31.0994
512	0.0122	86.6229
1024	0.0576	72.9906
1280	0.0942	69.7283
2048	0.2259	74.3916
4096	0.8418	79.7753

Nvidia Geforce Gtx 1650 — layout = row-major

n	compute ms	GB/s
32	0.0076	0.5858
64	0.0078	2.2058
128	0.0080	8.4346
256	0.0092	28.6782
512	0.0120	87.5432
1024	0.0369	114.1111
1280	0.0541	121.3956
2048	0.1206	139.3461
4096	0.4461	150.5569

See also:

• layout.ssyr2.js — WebGPU layout-sweep benchmark script