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

Benchmark results for dtrsv 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	0.0840	0.0564	0.0125	0.3785	14.9%
64	0.1668	0.1059	0.0170	1.0395	10.2%
128	0.2171	0.3137	0.0257	2.6534	11.8%
256	0.3803	0.7028	0.0420	6.3610	11.0%
512	0.7778	1.3614	0.0783	13.5302	10.1%
1024	1.5734	2.6788	0.1512	27.8844	9.6%
2048	3.4299	4.9034	0.3105	54.1655	9.1%
4096	8.1634	8.2307	0.8373	80.2473	10.3%

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

See also

• dtrsv.js — WebGPU benchmark script
• dtrsv.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	0.0840	0.0564	0.0118	0.4000	14.1%
64	0.1667	0.1060	0.0156	1.1358	9.3%
128	0.2171	0.3137	0.0228	2.9846	10.5%
256	0.3813	0.7010	0.0380	7.0244	10.0%
512	0.7823	1.3534	0.0714	14.8377	9.1%
1024	1.5772	2.6723	0.1400	30.0953	8.9%
1280	2.0089	3.2750	0.1765	37.2835	8.8%
2048	3.4419	4.8863	0.2949	57.0216	8.6%
4096	8.2883	8.1067	0.6959	96.5496	8.4%

Nvidia Geforce Gtx 1650 — stride = 5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0122	0.3895	14.5%
64	0.1666	0.1060	0.0151	1.1732	9.0%
128	0.2176	0.3130	0.0226	3.0184	10.4%
256	0.3831	0.6977	0.0380	7.0422	9.9%
512	0.7820	1.3540	0.0710	14.9045	9.1%
1024	1.5705	2.6838	0.1393	30.2647	8.9%
1280	2.0007	3.2884	0.1758	37.4329	8.8%
2048	3.4160	4.9234	0.2953	56.9474	8.6%
4096	8.1674	8.2267	0.6944	96.7587	8.5%

Nvidia Geforce Gtx 1650 — stride = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0109	0.4327	13.0%
64	0.1669	0.1058	0.0147	1.2052	8.8%
128	0.2187	0.3113	0.0225	3.0227	10.3%
256	0.3856	0.6931	0.0378	7.0660	9.8%
512	0.7971	1.3284	0.0747	14.1796	9.4%
1024	1.6468	2.5593	0.1412	29.8396	8.6%
1280	2.1240	3.0975	0.1777	37.0184	8.4%
2048	3.7376	4.4997	0.2961	56.7967	7.9%
4096	9.4964	7.0754	0.6834	98.3149	7.2%

Nvidia Geforce Gtx 1650 — stride = 33

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0110	0.4296	13.1%
64	0.1667	0.1059	0.0148	1.1948	8.9%
128	0.2176	0.3130	0.0225	3.0227	10.4%
256	0.3819	0.6998	0.0393	6.8013	10.3%
512	0.7844	1.3498	0.0738	14.3518	9.4%
1024	1.5790	2.6692	0.1452	29.0242	9.2%
1280	2.0084	3.2758	0.1771	37.1454	8.8%
2048	3.4354	4.8955	0.2948	57.0494	8.6%
4096	8.2270	8.1671	0.6837	98.2689	8.3%

Nvidia Geforce Gtx 1650 — stride = 255

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0110	0.4302	13.1%
64	0.1674	0.1055	0.0219	0.8070	13.1%
128	0.2175	0.3130	0.0245	2.7817	11.3%
256	0.3826	0.6986	0.0444	6.0238	11.6%
512	0.7851	1.3487	0.0781	13.5551	9.9%
1024	1.5809	2.6661	0.1447	29.1302	9.2%
1280	2.0078	3.2768	0.1807	36.4055	9.0%
2048	3.4363	4.8942	0.2931	57.3733	8.5%
4096	8.2337	8.1605	0.6831	98.3656	8.3%

Nvidia Geforce Gtx 1650 — stride = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0108	0.4366	12.9%
64	0.1675	0.1055	0.0215	0.8202	12.9%
128	0.2190	0.3110	0.0240	2.8336	11.0%
256	0.3844	0.6953	0.0442	6.0434	11.5%
512	0.7987	1.3256	0.0783	13.5163	9.8%
1024	1.6500	2.5543	0.1438	29.3117	8.7%
1280	2.1215	3.1011	0.1796	36.6292	8.5%
2048	3.7438	4.4923	0.2951	56.9845	7.9%
4096	9.4956	7.0760	0.6829	98.3932	7.2%

See also:

• stride.dtrsv.js — WebGPU stride-sweep benchmark script
• stride.dtrsv.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	0.0840	0.0564	0.0119	0.3995	14.1%
64	0.1666	0.1060	0.0157	1.1242	9.4%
128	0.2171	0.3137	0.0231	2.9453	10.7%
256	0.3807	0.7021	0.0389	6.8684	10.2%
512	0.7828	1.3525	0.0713	14.8410	9.1%
1024	1.5712	2.6825	0.1356	31.0935	8.6%
1280	1.9987	3.2918	0.1700	38.7121	8.5%
2048	3.4119	4.9292	0.2804	59.9895	8.2%
4096	8.1541	8.2401	0.6648	101.0765	8.2%

Nvidia Geforce Gtx 1650 — trans = transpose

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0858	0.0552	0.0191	0.2479	22.3%
64	0.1720	0.1027	0.0228	0.7753	13.2%
128	0.2394	0.2845	0.0350	1.9452	14.6%
256	0.4244	0.6297	0.0593	4.5036	14.0%
512	0.8481	1.2484	0.1126	9.4000	13.3%
1024	1.7649	2.3881	0.2249	18.7397	12.7%
1280	2.2839	2.8807	0.2848	23.0985	12.5%
2048	3.9621	4.2447	0.4649	36.1762	11.7%
4096	9.9004	6.7867	1.0240	65.6160	10.3%

See also:

• trans.dtrsv.js — WebGPU trans-sweep benchmark script
• trans.dtrsv.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	0.0840	0.0564	0.0120	0.3957	14.3%
64	0.1674	0.1055	0.0155	1.1405	9.3%
128	0.2172	0.3135	0.0235	2.9031	10.8%
256	0.3802	0.7029	0.0384	6.9542	10.1%
512	0.7787	1.3598	0.0715	14.8144	9.2%
1024	1.5711	2.6826	0.1358	31.0422	8.6%
1280	1.9968	3.2949	0.1702	38.6575	8.5%
2048	3.4117	4.9296	0.2803	60.0100	8.2%
4096	8.1575	8.2367	0.6642	101.1641	8.1%

Nvidia Geforce Gtx 1650 — uplo = upper

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0847	0.0559	0.0118	0.4005	14.0%
64	0.1679	0.1052	0.0155	1.1405	9.2%
128	0.2218	0.3070	0.0227	3.0056	10.2%
256	0.3994	0.6691	0.0382	6.9979	9.6%
512	0.7782	1.3607	0.0701	15.0983	9.0%
1024	1.5753	2.6756	0.1356	31.0935	8.6%
1280	2.0038	3.2833	0.1700	38.7048	8.5%
2048	3.4120	4.9292	0.2792	60.2439	8.2%
4096	8.1572	8.2370	0.6595	101.8882	8.1%

See also:

• uplo.dtrsv.js — WebGPU uplo-sweep benchmark script
• uplo.dtrsv.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	0.0840	0.0564	0.0117	0.4055	13.9%
64	0.1666	0.1060	0.0156	1.1358	9.3%
128	0.2171	0.3137	0.0246	2.7708	11.3%
256	0.3808	0.7018	0.0384	6.9629	10.1%
512	0.7773	1.3622	0.0710	14.9045	9.1%
1024	1.5684	2.6872	0.1353	31.1597	8.6%
1280	2.0009	3.2882	0.1706	38.5669	8.5%
2048	3.4138	4.9265	0.2803	59.9929	8.2%
4096	8.1559	8.2383	0.6640	101.1958	8.1%

Nvidia Geforce Gtx 1650 — pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0111	0.4265	13.2%
64	0.1679	0.1052	0.0148	1.1897	8.8%
128	0.2175	0.3131	0.0225	3.0227	10.4%
256	0.3816	0.7003	0.0376	7.1020	9.9%
512	0.7823	1.3534	0.0681	15.5379	8.7%
1024	1.5688	2.6867	0.1352	31.1818	8.6%
1280	1.9981	3.2927	0.1700	38.6939	8.5%
2048	3.4130	4.9277	0.2888	58.2379	8.5%

Nvidia Geforce Gtx 1650 — pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0118	0.4011	14.1%
64	0.1667	0.1060	0.0155	1.1417	9.3%
128	0.2170	0.3139	0.0230	2.9556	10.6%
256	0.3812	0.7011	0.0382	7.0038	10.0%
512	0.7805	1.3565	0.0695	15.2339	8.9%
1024	1.5708	2.6832	0.1353	31.1450	8.6%
1280	2.0012	3.2876	0.1700	38.7048	8.5%
2048	3.4099	4.9322	0.2853	58.9400	8.4%

Nvidia Geforce Gtx 1650 — pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0117	0.4044	13.9%
64	0.1666	0.1060	0.0156	1.1358	9.3%
128	0.2171	0.3137	0.0228	2.9846	10.5%
256	0.3809	0.7016	0.0395	6.7600	10.4%
512	0.7837	1.3510	0.0708	14.9483	9.0%
1024	1.5690	2.6864	0.1372	30.7164	8.7%
1280	1.9948	3.2982	0.1717	38.3225	8.6%
2048	3.4126	4.9282	0.2817	59.7067	8.3%

Nvidia Geforce Gtx 1650 — pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0117	0.4033	14.0%
64	0.1668	0.1059	0.0156	1.1323	9.4%
128	0.2172	0.3135	0.0229	2.9762	10.5%
256	0.3814	0.7007	0.0384	6.9658	10.1%
512	0.7823	1.3534	0.0701	15.0949	9.0%
1024	1.5730	2.6795	0.1366	30.8567	8.7%
1280	1.9996	3.2903	0.1700	38.7048	8.5%
2048	3.4121	4.9289	0.2821	59.6254	8.3%

Nvidia Geforce Gtx 1650 — pad = 48

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0118	0.4022	14.0%
64	0.1674	0.1055	0.0154	1.1488	9.2%
128	0.2171	0.3137	0.0229	2.9762	10.5%
256	0.3802	0.7030	0.0387	6.9025	10.2%
512	0.7824	1.3534	0.0696	15.2059	8.9%
1024	1.5681	2.6878	0.1358	31.0312	8.7%
1280	1.9977	3.2934	0.1700	38.7048	8.5%
2048	3.4121	4.9290	0.2814	59.7712	8.2%

Nvidia Geforce Gtx 1650 — pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0118	0.4000	14.1%
64	0.1675	0.1054	0.0155	1.1417	9.2%
128	0.2171	0.3137	0.0230	2.9556	10.6%
256	0.3808	0.7018	0.0385	6.9398	10.1%
512	0.7802	1.3572	0.0699	15.1571	9.0%
1024	1.5697	2.6851	0.1372	30.7164	8.7%
1280	1.9988	3.2915	0.1703	38.6248	8.5%
2048	3.4125	4.9284	0.2806	59.9416	8.2%

Nvidia Geforce Gtx 1650 — pad = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0840	0.0564	0.0116	0.4088	13.8%
64	0.1674	0.1055	0.0155	1.1417	9.2%
128	0.2174	0.3133	0.0231	2.9474	10.6%
256	0.3809	0.7017	0.0383	6.9774	10.1%
512	0.7810	1.3557	0.0704	15.0332	9.0%
1024	1.5697	2.6851	0.1352	31.1818	8.6%
1280	1.9968	3.2949	0.1700	38.7048	8.5%
2048	3.4099	4.9322	0.2824	59.5646	8.3%

See also:

• lda.dtrsv.js — WebGPU lda-sweep benchmark script
• lda.dtrsv.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	0.1073	0.0441	0.0163	0.2913	15.2%
64	0.2139	0.0826	0.0208	0.8486	9.7%
128	0.2787	0.2443	0.0299	2.2784	10.7%
256	0.4882	0.5474	0.0509	5.2462	10.4%
512	0.9949	1.0642	0.0903	11.7313	9.1%
1024	1.8841	2.2370	0.1732	24.3325	9.2%
1280	1.9973	3.2941	0.2172	30.2932	10.9%
2048	3.4119	4.9293	0.3597	46.7566	10.5%
4096	8.1596	8.2346	0.8315	80.8048	10.2%

Nvidia Geforce Gtx 1650 — diag = unit

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0734	0.0645	0.0081	0.5827	11.1%
64	0.1451	0.1217	0.0111	1.5977	7.6%
128	0.1910	0.3565	0.0177	3.8446	9.3%
256	0.3372	0.7927	0.0307	8.7000	9.1%
512	0.6940	1.5256	0.0576	18.3976	8.3%
1024	1.4174	2.9737	0.1132	37.2383	8.0%
1280	1.8103	3.6344	0.1413	46.5474	7.8%
2048	3.1153	5.3985	0.2336	72.0104	7.5%
4096	7.5748	8.8703	0.5390	124.6527	7.1%

See also:

• diag.dtrsv.js — WebGPU diag-sweep benchmark script
• diag.dtrsv.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	0.0853	0.0555
64	0.1741	0.1015
128	0.2331	0.2921
256	0.4137	0.6460
512	0.8628	1.2271
1024	1.7654	2.3874
1280	2.2853	2.8789
2048	3.9579	4.2492
4096	9.9104	6.7798

Nvidia Geforce Gtx 1650 — layout = row-major

n	compute ms	GB/s
32	0.0840	0.0564
64	0.1665	0.1061
128	0.2173	0.3133
256	0.3805	0.7024
512	0.7800	1.3575
1024	1.5684	2.6874
1280	1.9990	3.2912
2048	3.4120	4.9291
4096	8.1567	8.2375

See also:

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