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

Benchmark results for ssyr 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	0.5726	0.0040	1.0924	52.4%
64	0.0079	2.1507	0.0038	4.4746	48.1%
128	0.0079	8.3871	0.0041	16.3780	51.2%
256	0.0086	30.6914	0.0038	69.3782	44.2%
512	0.0118	89.1491	0.0061	172.2304	51.8%
1024	0.0553	76.0000	0.0312	134.8337	56.4%
1280	0.0942	69.6739	0.0686	95.6270	72.9%
2048	0.2191	76.6355	0.1188	141.3793	54.2%
4096	0.8264	81.2459	0.4532	148.1344	54.8%

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

See also

• ssyr.js — WebGPU benchmark script
• ssyr.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	0.0077	0.5667	0.0041	1.0502	54.0%
64	0.0079	2.1420	0.0040	4.1905	51.1%
128	0.0082	8.1250	0.0040	16.7742	48.4%
256	0.0092	28.6667	0.0044	60.4835	47.4%
512	0.0122	86.1152	0.0067	157.0215	54.8%
1024	0.0369	114.0000	0.0314	133.6672	85.3%
1280	0.0544	120.6588	0.0676	97.1672	124.2%
2048	0.1218	137.8694	0.1187	141.4555	97.5%
4096	0.4608	145.7067	0.4648	144.4428	100.9%

Nvidia Geforce Gtx 1650 — stride = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0079	0.5528	0.0040	1.0880	50.8%
64	0.0078	2.1639	0.0041	4.1575	52.0%
128	0.0083	8.0464	0.0041	16.2500	49.5%
256	0.0122	21.6693	0.0041	64.0000	33.9%
512	0.0247	42.6944	0.0074	141.7931	30.1%
1024	0.0891	47.1470	0.0322	130.3504	36.2%
1280	0.1372	47.8358	0.0676	97.1212	49.3%
2048	0.3450	48.6827	0.1207	139.0936	35.0%
4096	1.3312	50.4369	0.4854	138.3291	36.5%

Nvidia Geforce Gtx 1650 — stride = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0063	0.6939	0.0041	1.0625	65.3%
64	0.0062	2.7287	0.0041	4.1250	66.1%
128	0.0070	9.4977	0.0043	15.5224	61.2%
256	0.0098	26.8925	0.0042	63.5077	42.3%
512	0.0192	54.7354	0.0070	149.8679	36.5%
1024	0.0739	56.8643	0.0325	129.2598	44.0%
1280	0.1160	56.5926	0.0665	98.7578	57.3%
2048	0.2911	57.6925	0.1172	143.2510	40.3%
4096	1.1339	59.2145	0.4772	140.7086	42.1%

See also:

• stride.ssyr.js — WebGPU stride-sweep benchmark script
• stride.ssyr.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	0.7083	0.0041	1.0584	66.9%
64	0.0061	2.7500	0.0039	4.3636	63.0%
128	0.0063	10.5051	0.0040	16.5079	63.6%
256	0.0076	34.7621	0.0041	64.5000	53.9%
512	0.0096	109.2890	0.0102	102.8000	106.3%
1024	0.0331	127.0097	0.0313	134.3509	94.5%
1280	0.0512	128.2802	0.0687	95.5157	134.3%
2048	0.1147	146.4286	0.1165	144.0966	101.6%
4096	0.4340	154.7099	0.4641	144.6818	106.9%

Nvidia Geforce Gtx 1650 — uplo = upper

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0073	0.5939	0.0041	1.0625	55.9%
64	0.0074	2.2957	0.0040	4.1905	54.8%
128	0.0076	8.7579	0.0040	16.5737	52.8%
256	0.0085	31.0376	0.0041	63.7529	48.7%
512	0.0119	88.7881	0.0068	154.4413	57.5%
1024	0.0570	73.7591	0.0326	128.7529	57.3%
1280	0.0913	71.8585	0.0673	97.5368	73.7%
2048	0.2122	79.1315	0.1171	143.4468	55.2%
4096	0.8044	83.4680	0.4570	146.9157	56.8%

See also:

• uplo.ssyr.js — WebGPU uplo-sweep benchmark script
• uplo.ssyr.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	0.0067	0.6507	0.0043	1.0226	63.6%
64	0.0068	2.4965	0.0041	4.0772	61.2%
128	0.0071	9.3483	0.0040	16.6400	56.2%
256	0.0082	32.2500	0.0041	64.0000	50.4%
512	0.0104	101.2185	0.0067	156.6476	64.6%
1024	0.0348	120.7059	0.0314	133.8716	90.2%
1280	0.0522	125.6863	0.0683	96.1650	130.7%
2048	0.1167	143.8596	0.1188	141.3793	101.8%
4096	0.4360	154.0059	0.4644	144.5772	106.5%

Nvidia Geforce Gtx 1650 — pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0068	0.6415	0.0040	1.0837	59.2%
64	0.0069	2.4615	0.0040	4.2581	57.8%
128	0.0075	8.8511	0.0040	16.7742	52.8%
256	0.0083	32.0000	0.0050	52.4190	61.0%
512	0.0112	93.9886	0.0136	77.4024	121.4%
1024	0.0573	73.2857	0.0409	102.7606	71.3%
1280	0.0902	72.7763	0.0770	85.2890	85.3%
2048	0.2151	78.0836	0.1438	116.7649	66.9%
4096	0.8132	82.5680	0.7674	87.4933	94.4%

Nvidia Geforce Gtx 1650 — pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7083	0.0040	1.0794	65.6%
64	0.0062	2.7358	0.0039	4.2927	63.7%
128	0.0065	10.2463	0.0038	17.3333	59.1%
256	0.0076	34.9091	0.0041	64.5000	54.1%
512	0.0096	109.2890	0.0067	156.2755	69.9%
1024	0.0549	76.4869	0.0435	96.5292	79.2%
1280	0.0900	72.9705	0.0756	86.8049	84.1%
2048	0.2015	83.3413	0.1809	92.8274	89.8%
4096	0.7525	89.2290	0.7659	87.6596	101.8%

Nvidia Geforce Gtx 1650 — pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7083	0.0035	1.2308	57.6%
64	0.0061	2.7500	0.0036	4.7143	58.3%
128	0.0065	10.1961	0.0036	18.5714	54.9%
256	0.0075	35.1319	0.0038	69.3782	50.6%
512	0.0096	110.2044	0.0061	171.3333	64.3%
1024	0.0549	76.4869	0.0443	94.9588	80.5%
1280	0.0860	76.3095	0.0702	93.5127	81.6%
2048	0.2064	81.3454	0.1446	116.1319	70.0%
4096	0.7559	88.8267	0.7598	88.3666	100.5%

Nvidia Geforce Gtx 1650 — pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7083	0.0047	0.9189	77.1%
64	0.0061	2.7500	0.0041	4.1250	66.7%
128	0.0065	10.2211	0.0040	16.6400	61.4%
256	0.0076	34.8354	0.0041	64.5000	54.0%
512	0.0096	109.2890	0.0067	157.3971	69.4%
1024	0.0341	123.1392	0.0389	108.0000	114.0%
1280	0.0512	128.2000	0.0595	110.3983	116.1%
2048	0.1147	146.3673	0.1364	123.0769	118.9%
4096	0.4468	150.2561	0.6035	111.2501	135.1%

Nvidia Geforce Gtx 1650 — pad = 48

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7083	0.0034	1.2651	56.0%
64	0.0061	2.7500	0.0034	4.9577	55.5%
128	0.0065	10.2716	0.0146	4.5564	225.4%
256	0.0076	34.5439	0.0038	69.3782	49.8%
512	0.0098	107.8557	0.0061	171.3333	63.0%
1024	0.0567	74.0711	0.0451	93.2727	79.4%
1280	0.0844	77.7411	0.0696	94.3731	82.4%
2048	0.2172	77.3129	0.1700	98.7952	78.3%
4096	0.8037	83.5378	0.7930	84.6653	98.7%

Nvidia Geforce Gtx 1650 — pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0066	0.6618	0.0034	1.2710	52.1%
64	0.0066	2.5507	0.0036	4.7568	53.6%
128	0.0072	9.2650	0.0035	19.2593	48.1%
256	0.0082	32.2500	0.0037	71.1724	45.3%
512	0.0104	101.6878	0.0061	173.1368	58.7%
1024	0.0362	115.9629	0.0413	101.8441	113.9%
1280	0.0523	125.4173	0.0568	115.5280	108.6%
2048	0.1226	136.9699	0.1468	114.4228	119.7%
4096	0.4506	149.0182	0.6023	111.4806	133.7%

Nvidia Geforce Gtx 1650 — pad = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0072	0.6018	0.0041	1.0625	56.6%
64	0.0075	2.2612	0.0041	4.1250	54.8%
128	0.0082	8.1250	0.0043	15.5224	52.3%
256	0.0091	29.0193	0.0040	66.3132	43.8%
512	0.0120	87.4894	0.0065	162.4494	53.9%
1024	0.0385	109.2579	0.0414	101.4900	107.7%
1280	0.0557	117.8512	0.0614	106.8612	110.3%
2048	0.1276	131.6278	0.1902	88.3056	149.1%
4096	0.5204	129.0113	0.6364	105.4969	122.3%

See also:

• lda.ssyr.js — WebGPU lda-sweep benchmark script
• lda.ssyr.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	0.0075	0.5775	0.0036	1.2197	47.3%
64	0.0077	2.1863	0.0035	4.8889	44.7%
128	0.0080	8.3534	0.0036	18.7387	44.6%
256	0.0090	29.2766	0.0038	69.9661	41.8%
512	0.0120	87.9572	0.0061	172.6824	50.9%
1024	0.0367	114.4969	0.0310	135.5294	84.5%
1280	0.0537	122.1316	0.0700	93.7477	130.3%
2048	0.1201	139.8162	0.1187	141.4555	98.8%
4096	0.4444	151.0783	0.4462	150.4716	100.4%

Nvidia Geforce Gtx 1650 — alpha = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7083	0.0017	2.5660	27.6%
64	0.0061	2.7500	0.0015	11.1158	24.7%
128	0.0063	10.5316	0.0015	43.3333	24.3%
256	0.0075	35.2068	0.0015	172.0000	20.5%
512	0.0096	109.6533	0.0016	657.9200	16.7%
1024	0.0334	125.7931	0.0014	3019.0342	4.2%
1280	0.0511	128.4408	0.0014	4715.4019	2.7%
2048	0.1147	146.4286	0.0015	11165.9570	1.3%
4096	0.4328	155.1332	0.0014	47685.8164	0.3%

Nvidia Geforce Gtx 1650 — alpha = 1e-38

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0074	0.5849	0.0035	1.2420	47.1%
64	0.0076	2.2232	0.0034	4.9346	45.1%
128	0.0080	8.3034	0.0035	18.9954	43.7%
256	0.0092	28.8671	0.0038	69.3782	41.6%
512	0.0119	88.6685	0.0061	171.3333	51.8%
1024	0.0412	101.8836	0.0310	135.5294	75.2%
1280	0.0600	109.3099	0.0699	93.9624	116.3%
2048	0.1388	120.9635	0.1192	140.8670	85.9%
4096	0.5202	129.0629	0.4463	150.4446	85.8%

Nvidia Geforce Gtx 1650 — alpha = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0061	0.7083	0.0041	1.0625	66.7%
64	0.0061	2.7500	0.0041	4.1250	66.7%
128	0.0063	10.5051	0.0041	16.2500	64.6%
256	0.0076	34.5439	0.0044	60.2628	57.3%
512	0.0098	107.6792	0.0068	155.9052	69.1%
1024	0.0334	125.6727	0.0414	101.5684	123.7%
1280	0.0512	128.2802	0.0711	92.2717	139.0%
2048	0.1147	146.4286	0.1154	145.4747	100.7%
4096	0.4440	151.2308	0.4560	147.2404	102.7%

Nvidia Geforce Gtx 1650 — alpha = 2.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0076	0.5751	0.0036	1.1982	48.0%
64	0.0076	2.2185	0.0035	4.8000	46.2%
128	0.0081	8.2376	0.0037	17.7778	46.3%
256	0.0094	28.0816	0.0039	67.9506	41.3%
512	0.0120	87.7227	0.0062	169.5670	51.7%
1024	0.0366	114.7471	0.0310	135.5294	84.7%
1280	0.0543	120.9434	0.0655	100.1562	120.8%
2048	0.1203	139.5745	0.1133	148.1858	94.2%
4096	0.4444	151.0783	0.4452	150.8177	100.2%

See also:

• alpha.ssyr.js — WebGPU alpha-sweep benchmark script
• alpha.ssyr.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	0.5714
64	0.0078	2.1639
128	0.0080	8.3534
256	0.0086	30.5778
512	0.0119	88.7881
1024	0.0575	73.0818
1280	0.0963	68.1802
2048	0.2163	77.6274
4096	0.8280	81.0921

Nvidia Geforce Gtx 1650 — layout = row-major

n	compute ms	GB/s
32	0.0076	0.5702
64	0.0077	2.1818
128	0.0080	8.3200
256	0.0090	29.2248
512	0.0119	88.6685
1024	0.0368	114.3474
1280	0.0540	121.6247
2048	0.1199	140.0213
4096	0.4444	151.0783

See also:

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