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

Benchmark results for strmv 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	0.0082	0.2891	0.0048	0.4933	58.6%
64	0.0083	1.0698	0.0064	1.3731	77.9%
128	0.0092	3.7138	0.0099	3.4323	108.2%
256	0.0146	9.1679	0.0170	7.8570	116.7%
512	0.0270	19.6251	0.0348	15.1989	129.1%
1024	0.0772	27.3035	0.1718	12.2671	222.6%
1280	0.1206	27.2715	0.0662	49.6738	54.9%
2048	0.2635	31.9145	0.0996	84.4286	37.8%
4096	1.4785	22.7227	0.2533	132.6497	17.1%

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

strmv-default GB/s chart

strmv-default ms chart

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	0.0068	0.3507	0.0049	0.4805	73.0%
64	0.0068	1.2927	0.0066	1.3463	96.0%
128	0.0070	4.8364	0.0099	3.4323	140.9%
256	0.0086	15.5242	0.0169	7.8867	196.8%
512	0.0114	46.4067	0.0343	15.4545	300.3%
1024	0.0287	73.5000	0.1717	12.2763	598.7%
1280	0.0389	84.4700	0.0672	48.9757	172.5%
2048	0.0825	101.8938	0.1003	83.8226	121.6%
4096	0.2771	121.2235	0.2540	132.2903	91.6%

strmv-transno-transpose GB/s chart

strmv-transno-transpose ms chart

Nvidia Geforce Gtx 1650 — trans = transpose

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0068	0.3458	0.0051	0.4625	74.8%
64	0.0070	1.2632	0.0067	1.3206	95.7%
128	0.0080	4.2560	0.0123	2.7708	153.6%
256	0.0116	11.5359	0.0321	4.1614	277.2%
512	0.0208	25.4523	0.1142	4.6342	549.2%
1024	0.0608	34.6884	0.4506	4.6773	741.6%
1280	0.0977	33.6608	0.2152	15.2828	220.3%
2048	0.2134	39.4097	0.3280	25.6400	153.7%
4096	1.1402	29.4634	1.0297	32.6270	90.3%

strmv-transtranspose GB/s chart

strmv-transtranspose ms chart

See also:

trans.strmv.js — WebGPU trans-sweep benchmark script
trans.strmv.c — CUDA / cuBLAS trans-sweep reference script

Lda sweep

Unless noted otherwise, every result above uses a tight lda (no padding). Padding the row stride only matters for trans = "transpose" here (swept at both trans values below so that's visible in the data, not just claimed). Collapsed below by default — expand a trans value, then a pad, to see its table and chart.

Nvidia Geforce Gtx 1650 — trans = no-transpose (8 pads)

pad = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.2879	0.0049	0.4868	59.1%
64	0.0083	1.0615	0.0062	1.4154	75.0%
128	0.0085	3.9925	0.0098	3.4658	115.2%
256	0.0106	12.5783	0.0168	7.9695	157.8%
512	0.0149	35.4641	0.0339	15.6002	227.3%
1024	0.0338	62.3636	0.1707	12.3488	505.0%
1280	0.0451	72.9337	0.0673	48.8709	149.2%
2048	0.0942	89.2609	0.0988	85.0984	104.9%
4096	0.3119	107.7107	0.2532	132.6916	81.2%

strmv-lda-no-transpose-pad0 GB/s chart

strmv-lda-no-transpose-pad0 ms chart

pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.2891	0.0049	0.4868	59.4%
64	0.0082	1.0739	0.0064	1.3869	77.4%
128	0.0087	3.9190	0.0100	3.4157	114.7%
256	0.0107	12.4843	0.0167	7.9847	156.4%
512	0.0147	35.9652	0.0339	15.6223	230.2%
1024	0.0344	61.3184	0.1696	12.4233	493.6%
1280	0.0455	72.2417	0.0676	48.6742	148.4%
2048	0.0951	88.4348	0.1008	83.4235	106.0%
4096	0.3174	105.8323	0.2598	129.3245	81.8%

strmv-lda-no-transpose-pad1 GB/s chart

strmv-lda-no-transpose-pad1 ms chart

pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.2891	0.0049	0.4868	59.4%
64	0.0082	1.0760	0.0064	1.3766	78.2%
128	0.0086	3.9407	0.0099	3.4323	114.8%
256	0.0107	12.5405	0.0166	8.0462	155.9%
512	0.0149	35.5403	0.0339	15.6223	227.5%
1024	0.0345	61.1476	0.1700	12.3999	493.1%
1280	0.0464	70.9455	0.0667	49.3046	143.9%
2048	0.0922	91.1811	0.0961	87.4926	104.2%
4096	0.3033	110.7794	0.2492	134.8303	82.2%

strmv-lda-no-transpose-pad8 GB/s chart

strmv-lda-no-transpose-pad8 ms chart

pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.2891	0.0045	0.5230	55.3%
64	0.0082	1.0739	0.0061	1.4375	74.7%
128	0.0087	3.9190	0.0094	3.6376	107.7%
256	0.0107	12.5030	0.0158	8.4620	147.8%
512	0.0149	35.6168	0.0314	16.8644	211.2%
1024	0.0344	61.2614	0.1616	13.0369	469.9%
1280	0.0460	71.5130	0.0632	52.0111	137.5%
2048	0.0957	87.8289	0.0954	88.1677	99.6%
4096	0.3023	111.1194	0.2489	134.9776	82.3%

strmv-lda-no-transpose-pad16 GB/s chart

strmv-lda-no-transpose-pad16 ms chart

pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.2891	0.0047	0.5034	57.4%
64	0.0082	1.0781	0.0061	1.4375	75.0%
128	0.0086	3.9407	0.0091	3.7268	105.7%
256	0.0106	12.6163	0.0159	8.4194	149.8%
512	0.0148	35.7322	0.0316	16.7619	213.2%
1024	0.0344	61.2045	0.1618	13.0253	469.9%
1280	0.0452	72.7016	0.0635	51.8145	140.3%
2048	0.1162	72.3723	0.0953	88.1973	82.1%
4096	0.3829	87.7293	0.2494	134.6833	65.1%

strmv-lda-no-transpose-pad32 GB/s chart

strmv-lda-no-transpose-pad32 ms chart

pad = 48

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0118	0.2008	0.0046	0.5139	39.1%
64	0.0118	0.7500	0.0061	1.4375	52.2%
128	0.0123	2.7781	0.0092	3.6944	75.2%
256	0.0162	8.2693	0.0160	8.3687	98.8%
512	0.0205	25.8500	0.0315	16.8301	153.6%
1024	0.0453	46.5249	0.1618	13.0253	357.2%
1280	0.0601	54.7537	0.0635	51.8145	105.7%
2048	0.1208	69.5932	0.0961	87.5363	79.5%
4096	0.4020	83.5773	0.2505	134.1069	62.3%

strmv-lda-no-transpose-pad48 GB/s chart

strmv-lda-no-transpose-pad48 ms chart

pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0123	0.1932	0.0045	0.5230	36.9%
64	0.0123	0.7188	0.0061	1.4450	49.7%
128	0.0129	2.6402	0.0094	3.6068	73.2%
256	0.0168	7.9392	0.0158	8.4792	93.6%
512	0.0220	24.1166	0.0315	16.8045	143.5%
1024	0.0493	42.7636	0.1618	13.0253	328.3%
1280	0.0655	50.2198	0.0635	51.8145	96.9%
2048	0.1340	62.7694	0.0956	87.9906	71.3%
4096	0.3167	106.0782	0.2585	129.9648	81.6%

strmv-lda-no-transpose-pad64 GB/s chart

strmv-lda-no-transpose-pad64 ms chart

pad = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.2891	0.0048	0.4966	58.2%
64	0.0082	1.0781	0.0063	1.3939	77.3%
128	0.0085	4.0151	0.0099	3.4323	117.0%
256	0.0103	12.9689	0.0166	8.0308	161.5%
512	0.0144	36.8053	0.0335	15.8240	232.6%
1024	0.0331	63.6290	0.1685	12.5035	508.9%
1280	0.0448	73.5073	0.0660	49.8424	147.5%
2048	0.0928	90.6308	0.1008	83.3970	108.7%
4096	0.3296	101.9228	0.2601	129.1653	78.9%

strmv-lda-no-transpose-pad128 GB/s chart

strmv-lda-no-transpose-pad128 ms chart

Nvidia Geforce Gtx 1650 — trans = transpose (8 pads)

pad = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.2891	0.0052	0.4512	64.1%
64	0.0083	1.0698	0.0068	1.2897	82.9%
128	0.0092	3.7073	0.0123	2.7708	133.8%
256	0.0143	9.3214	0.0316	4.2224	220.8%
512	0.0259	20.4121	0.1127	4.6980	434.5%
1024	0.0742	28.3862	0.4464	4.7204	601.4%
1280	0.1175	27.9918	0.2744	11.9897	233.5%
2048	0.2632	31.9494	0.3338	25.1890	126.8%
4096	1.4029	23.9474	1.0420	32.2403	74.3%

strmv-lda-transpose-pad0 GB/s chart

strmv-lda-transpose-pad0 ms chart

pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0069	0.3434	0.0041	0.5781	59.4%
64	0.0069	1.2748	0.0055	1.6047	79.4%
128	0.0079	4.3077	0.0082	4.1562	103.6%
256	0.0098	13.6918	0.0162	8.2448	166.1%
512	0.0159	33.3548	0.0492	10.7708	309.7%
1024	0.0393	53.6505	0.1981	10.6391	504.3%
1280	0.0659	49.9393	0.1518	21.6740	230.4%
2048	0.1188	70.7931	0.1779	47.2803	149.7%
4096	0.8644	38.8648	0.5319	63.1663	61.5%

strmv-lda-transpose-pad1 GB/s chart

strmv-lda-transpose-pad1 ms chart

pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0068	0.3491	0.0041	0.5781	60.4%
64	0.0070	1.2632	0.0051	1.7358	72.8%
128	0.0078	4.3429	0.0078	4.3607	99.6%
256	0.0102	13.1321	0.0166	8.0385	163.4%
512	0.0175	30.2450	0.0567	9.3337	324.0%
1024	0.0451	46.7727	0.2314	9.1062	513.6%
1280	0.0700	46.9728	0.1467	22.4258	209.5%
2048	0.1475	57.0278	0.2191	38.3738	148.6%
4096	1.0693	31.4178	0.6512	51.5862	60.9%

strmv-lda-transpose-pad8 GB/s chart

strmv-lda-transpose-pad8 ms chart

pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0068	0.3491	0.0041	0.5781	60.4%
64	0.0069	1.2837	0.0054	1.6235	79.1%
128	0.0079	4.3340	0.0084	4.0533	106.9%
256	0.0105	12.7317	0.0208	6.4395	197.7%
512	0.0184	28.7222	0.0675	7.8408	366.3%
1024	0.0512	41.1600	0.2827	7.4557	552.1%
1280	0.0758	43.4122	0.1558	21.1088	205.7%
2048	0.1768	47.5541	0.2498	33.6579	141.3%
4096	1.1587	28.9939	0.7482	44.9031	64.6%

strmv-lda-transpose-pad16 GB/s chart

strmv-lda-transpose-pad16 ms chart

pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0068	0.3474	0.0042	0.5606	62.0%
64	0.0068	1.2897	0.0056	1.5771	81.8%
128	0.0081	4.1890	0.0101	3.3778	124.0%
256	0.0114	11.6975	0.0257	5.2005	224.9%
512	0.0216	24.5278	0.0921	5.7464	426.8%
1024	0.0616	34.2287	0.3579	5.8879	581.3%
1280	0.0917	35.8751	0.1985	16.5753	216.4%
2048	0.2253	37.3273	0.3581	23.4849	158.9%
4096	1.1121	30.2099	1.0467	32.0978	94.1%

strmv-lda-transpose-pad32 GB/s chart

strmv-lda-transpose-pad32 ms chart

pad = 48

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0068	0.3507	0.0041	0.5781	60.7%
64	0.0070	1.2661	0.0053	1.6527	76.6%
128	0.0079	4.3252	0.0085	3.9850	108.5%
256	0.0106	12.6545	0.0205	6.5250	193.9%
512	0.0193	27.4362	0.0720	7.3529	373.1%
1024	0.0516	40.8156	0.2832	7.4405	548.6%
1280	0.0756	43.5224	0.1567	20.9967	207.3%
2048	0.1886	44.5812	0.2784	30.2016	147.6%
4096	1.1215	29.9569	0.7864	42.7187	70.1%

strmv-lda-transpose-pad48 GB/s chart

strmv-lda-transpose-pad48 ms chart

pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0069	0.3442	0.0042	0.5692	60.5%
64	0.0072	1.2321	0.0056	1.5862	77.7%
128	0.0080	4.2560	0.0098	3.4602	123.0%
256	0.0116	11.5679	0.0268	4.9833	232.1%
512	0.0212	24.9157	0.0901	5.8750	424.1%
1024	0.0635	33.1935	0.3758	5.6079	591.9%
1280	0.0924	35.5832	0.2036	16.1534	220.3%
2048	0.2216	37.9554	0.3441	24.4371	155.3%
4096	1.1059	30.3778	1.0835	31.0053	98.0%

strmv-lda-transpose-pad64 GB/s chart

strmv-lda-transpose-pad64 ms chart

pad = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0068	0.3466	0.0042	0.5649	61.4%
64	0.0071	1.2377	0.0057	1.5376	80.5%
128	0.0079	4.2903	0.0099	3.4267	125.2%
256	0.0114	11.7303	0.0256	5.2265	224.4%
512	0.0216	24.4915	0.0956	5.5387	442.2%
1024	0.0624	33.7723	0.3606	5.8435	577.9%
1280	0.0908	36.2355	0.1949	16.8774	214.7%
2048	0.2335	36.0200	0.3618	23.2429	155.0%
4096	1.1469	29.2929	1.0461	32.1160	91.2%

strmv-lda-transpose-pad128 GB/s chart

strmv-lda-transpose-pad128 ms chart

See also:

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

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

Nvidia Geforce Gtx 1650 — wgblas vs cuBLAS

See also

Transpose sweep

Lda sweep

Settings

On This Page