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

Benchmark results for strsv 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	0.0369	0.0642	0.0084	0.2830	22.7%
64	0.0676	0.1307	0.0102	0.8625	15.2%
128	0.0834	0.4084	0.0140	2.4376	16.8%
256	0.1322	1.0111	0.0210	6.3756	15.9%
512	0.2736	1.9353	0.0354	14.9449	12.9%
1024	0.5813	3.6252	0.0828	25.4614	14.2%
2048	1.3230	6.3560	0.1697	49.5398	12.8%
4096	3.3645	9.9852	0.3760	89.3380	11.2%

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

See also

• strsv.js — WebGPU benchmark script
• strsv.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	0.0358	0.0661	0.0086	0.2761	24.0%
64	0.0649	0.1361	0.0104	0.8492	16.0%
128	0.0773	0.4407	0.0140	2.4348	18.1%
256	0.1162	1.1503	0.0208	6.4395	17.9%
512	0.2353	2.2500	0.0351	15.0949	14.9%
1024	0.4894	4.3064	0.0821	25.6598	16.8%
1280	0.6243	5.2696	0.1036	31.7578	16.6%
2048	1.0708	7.8533	0.1682	50.0065	15.7%
4096	2.8397	11.8307	0.3791	88.6179	13.4%

Nvidia Geforce Gtx 1650 — stride = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0358	0.0662	0.0082	0.2891	22.9%
64	0.0651	0.1356	0.0098	0.8990	15.1%
128	0.0776	0.4389	0.0134	2.5394	17.3%
256	0.1182	1.1303	0.0206	6.4744	17.5%
512	0.2418	2.1895	0.0355	14.9045	14.7%
1024	0.5183	4.0657	0.0864	24.4001	16.7%
1280	0.6730	4.8880	0.1105	29.7669	16.4%
2048	1.2017	6.9976	0.1756	47.8834	14.6%
4096	3.1294	10.7355	0.3911	85.8918	12.5%

Nvidia Geforce Gtx 1650 — stride = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0355	0.0667	0.0079	0.3008	22.2%
64	0.0652	0.1355	0.0098	0.9020	15.0%
128	0.0778	0.4379	0.0200	1.7038	25.7%
256	0.1182	1.1303	0.0220	6.0698	18.6%
512	0.2440	2.1697	0.0456	11.6098	18.7%
1024	0.5196	4.0562	0.0880	23.9476	16.9%
1280	0.6738	4.8822	0.1107	29.7239	16.4%
2048	1.2003	7.0059	0.1738	48.3859	14.5%
4096	3.1325	10.7247	0.3917	85.7690	12.5%

See also:

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

Transpose sweep

Unless noted otherwise, every result above uses trans = "no-transpose". trans = "transpose" reads A with a cross-thread lda-strided mirror pattern instead of a coalesced one, and the gap grows with n — collapsed below by default, expand a trans value to see its table and chart.

Nvidia Geforce Gtx 1650 — trans = no-transpose

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0356	0.0666	0.0092	0.2587	25.7%
64	0.0650	0.1359	0.0106	0.8313	16.3%
128	0.0774	0.4401	0.0142	2.3964	18.4%
256	0.1160	1.1525	0.0210	6.3513	18.1%
512	0.2314	2.2876	0.0354	14.9719	15.3%
1024	0.4874	4.3237	0.0833	25.3000	17.1%
1280	0.6226	5.2837	0.1044	31.4951	16.8%
2048	1.0650	7.8960	0.1696	49.5912	15.9%
4096	2.6925	12.4773	0.3768	89.1711	14.0%

Nvidia Geforce Gtx 1650 — trans = transpose

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0374	0.0633	0.0100	0.2357	26.9%
64	0.0671	0.1316	0.0119	0.7449	17.7%
128	0.0856	0.3980	0.0153	2.2259	17.9%
256	0.1415	0.9442	0.0226	5.9066	16.0%
512	0.2771	1.9107	0.0380	13.9142	13.7%
1024	0.5838	3.6096	0.0800	26.3266	13.7%
1280	0.7598	4.3295	0.1051	31.3081	13.8%
2048	1.3230	6.3560	0.1638	51.3501	12.4%
4096	3.3600	9.9985	0.4002	83.9515	11.9%

See also:

• trans.strsv.js — WebGPU trans-sweep benchmark script
• trans.strsv.c — CUDA / cuBLAS trans-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	0.0356	0.0666	0.0086	0.2751	24.2%
64	0.0653	0.1353	0.0106	0.8326	16.3%
128	0.0773	0.4405	0.0141	2.4100	18.3%
256	0.1150	1.1621	0.0213	6.2750	18.5%
512	0.2335	2.2675	0.0356	14.8844	15.2%
1024	0.4872	4.3252	0.0833	25.2903	17.1%
1280	0.6218	5.2908	0.1044	31.4951	16.8%
2048	1.0652	7.8943	0.1697	49.5492	15.9%
4096	2.8069	11.9689	0.3761	89.3266	13.4%

Nvidia Geforce Gtx 1650 — uplo = upper

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0366	0.0648	0.0084	0.2824	22.9%
64	0.0656	0.1345	0.0102	0.8625	15.6%
128	0.0787	0.4327	0.0138	2.4630	17.6%
256	0.1246	1.0723	0.0205	6.5250	16.4%
512	0.2353	2.2497	0.0350	15.1087	14.9%
1024	0.4896	4.3045	0.0671	31.4049	13.7%
1280	0.6256	5.2583	0.0835	39.4021	13.3%
2048	1.0670	7.8810	0.1302	64.5899	12.2%
4096	2.5761	13.0410	0.2888	116.3404	11.2%

See also:

• uplo.strsv.js — WebGPU uplo-sweep benchmark script
• uplo.strsv.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	0.0355	0.0667	0.0102	0.2313	28.8%
64	0.0653	0.1353	0.0125	0.7059	19.2%
128	0.0772	0.4413	0.0169	2.0190	21.9%
256	0.1159	1.1533	0.0265	5.0435	22.9%
512	0.2315	2.2867	0.0451	11.7500	19.5%
1024	0.4874	4.3235	0.0967	21.7850	19.8%
1280	0.6215	5.2930	0.1204	27.3114	19.4%
2048	1.0659	7.8892	0.1952	43.0829	18.3%
4096	2.6961	12.4609	0.4243	79.1836	15.7%

Nvidia Geforce Gtx 1650 — pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0369	0.0642	0.0099	0.2395	26.8%
64	0.0671	0.1317	0.0121	0.7282	18.1%
128	0.0790	0.4311	0.0164	2.0781	20.7%
256	0.1201	1.1124	0.0256	5.2200	21.3%
512	0.2412	2.1946	0.0444	11.9193	18.4%
1024	0.5038	4.1829	0.0949	22.2149	18.8%
1280	0.6427	5.1185	0.1188	27.6940	18.5%
2048	1.0979	7.6590	0.1802	46.6591	16.4%
4096	2.6166	12.8391	0.4055	82.8485	15.5%

Nvidia Geforce Gtx 1650 — pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0356	0.0666	0.0088	0.2686	24.8%
64	0.0650	0.1360	0.0112	0.7897	17.2%
128	0.0767	0.4440	0.0147	2.3206	19.1%
256	0.1163	1.1491	0.0230	5.8162	19.8%
512	0.2319	2.2833	0.0390	13.5829	16.8%
1024	0.4895	4.3054	0.0879	23.9825	18.0%
1280	0.6232	5.2784	0.1106	29.7454	17.7%
2048	1.0652	7.8941	0.1823	46.1348	17.1%
4096	2.5847	12.9980	0.4082	82.3094	15.8%

Nvidia Geforce Gtx 1650 — pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0356	0.0665	0.0092	0.2583	25.8%
64	0.0649	0.1360	0.0111	0.7988	17.0%
128	0.0766	0.4445	0.0150	2.2735	19.6%
256	0.1156	1.1565	0.0229	5.8324	19.8%
512	0.2315	2.2865	0.0391	13.5274	16.9%
1024	0.4875	4.3228	0.0886	23.7833	18.2%
1280	0.6226	5.2837	0.1106	29.7454	17.8%
2048	1.0669	7.8819	0.1819	46.2363	17.0%
4096	2.5743	13.0501	0.3932	85.4375	15.3%

Nvidia Geforce Gtx 1650 — pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0356	0.0666	0.0084	0.2835	23.5%
64	0.0648	0.1363	0.0102	0.8625	15.8%
128	0.0766	0.4446	0.0140	2.4348	18.3%
256	0.1157	1.1549	0.0209	6.3902	18.1%
512	0.2338	2.2648	0.0352	15.0400	15.1%
1024	0.4876	4.3218	0.0832	25.3292	17.1%
1280	0.6227	5.2826	0.1044	31.4951	16.8%
2048	1.0655	7.8922	0.1726	48.7314	16.2%
4096	2.5702	13.0709	0.3927	85.5593	15.3%

Nvidia Geforce Gtx 1650 — pad = 48

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0356	0.0665	0.0082	0.2891	23.0%
64	0.0648	0.1362	0.0102	0.8625	15.8%
128	0.0769	0.4427	0.0140	2.4376	18.2%
256	0.1152	1.1595	0.0210	6.3610	18.2%
512	0.2335	2.2675	0.0355	14.9314	15.2%
1024	0.4894	4.3064	0.0839	25.1311	17.1%
1280	0.6240	5.2722	0.1049	31.3462	16.8%
2048	1.0670	7.8810	0.1733	48.5199	16.2%
4096	2.5760	13.0415	0.3974	84.5329	15.4%

Nvidia Geforce Gtx 1650 — pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0357	0.0664	0.0082	0.2891	23.0%
64	0.0650	0.1359	0.0102	0.8625	15.8%
128	0.0770	0.4423	0.0140	2.4292	18.2%
256	0.1161	1.1512	0.0208	6.4197	17.9%
512	0.2332	2.2707	0.0353	14.9855	15.2%
1024	0.4875	4.3225	0.0831	25.3487	17.1%
1280	0.6244	5.2688	0.1044	31.4951	16.7%
2048	1.0650	7.8957	0.1736	48.4528	16.3%
4096	2.5635	13.1054	0.3953	84.9948	15.4%

Nvidia Geforce Gtx 1650 — pad = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0355	0.0667	0.0084	0.2835	23.5%
64	0.0651	0.1357	0.0102	0.8652	15.7%
128	0.0771	0.4418	0.0138	2.4687	17.9%
256	0.1155	1.1568	0.0206	6.4744	17.9%
512	0.2338	2.2648	0.0351	15.0949	15.0%
1024	0.4874	4.3234	0.0838	25.1407	17.2%
1280	0.6226	5.2837	0.1044	31.4951	16.8%
2048	1.0669	7.8816	0.1756	47.8965	16.5%
4096	2.5666	13.0897	0.4055	82.8485	15.8%

See also:

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

diag sweep

A unit diagonal lets the kernel skip the diagonal load — and for the triangular solve, the reciprocal as well — so any difference here is exactly that skipped work.

Nvidia Geforce Gtx 1650 — diag = non-unit

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0360	0.0658	0.0086	0.2756	23.9%
64	0.0651	0.1356	0.0107	0.8239	16.5%
128	0.0769	0.4428	0.0142	2.3991	18.5%
256	0.1155	1.1573	0.0209	6.3804	18.1%
512	0.2318	2.2843	0.0353	15.0127	15.2%
1024	0.4874	4.3235	0.0829	25.4074	17.0%
1280	0.6220	5.2887	0.1044	31.4951	16.8%
2048	1.0664	7.8855	0.1699	49.4885	15.9%
4096	2.7976	12.0086	0.3763	89.2887	13.4%

Nvidia Geforce Gtx 1650 — diag = unit

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0363	0.0652	0.0076	0.3103	21.0%
64	0.0659	0.1340	0.0086	1.0299	13.0%
128	0.0781	0.4362	0.0122	2.7890	15.6%
256	0.1188	1.1250	0.0187	7.1385	15.8%
512	0.2413	2.1936	0.0322	16.4617	13.3%
1024	0.5014	4.2034	0.0735	28.6829	14.7%
1280	0.6394	5.1445	0.0942	34.9244	14.7%
2048	1.0977	7.6604	0.1499	56.1085	13.7%
4096	2.5683	13.0808	0.3850	87.2553	15.0%

See also:

• diag.strsv.js — WebGPU diag-sweep benchmark script
• diag.strsv.c — CUDA / cuBLAS diag-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	0.0369	0.0641
64	0.0682	0.1295
128	0.0843	0.4041
256	0.1351	0.9893
512	0.2742	1.9307
1024	0.5870	3.5902
1280	0.7656	4.2966
2048	1.3302	6.3216
4096	3.3589	10.0020

Nvidia Geforce Gtx 1650 — layout = row-major

n	compute ms	GB/s
32	0.0356	0.0666
64	0.0650	0.1359
128	0.0772	0.4408
256	0.1161	1.1510
512	0.2333	2.2693
1024	0.4874	4.3235
1280	0.6226	5.2837
2048	1.0663	7.8864
4096	2.6699	12.5831

See also:

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