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

Benchmark results for sgemv 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	0.0136	0.3294	0.0053	0.8383	39.3%
64	0.0107	1.6000	0.0055	3.1345	51.0%
128	0.0148	4.5368	0.0058	11.4849	39.5%
256	0.0215	12.3333	0.0104	25.5802	48.2%
512	0.0502	21.0271	0.0142	74.0674	28.4%
1024	0.1085	38.7547	0.0328	128.3750	30.2%
1280	0.1372	47.8731	0.0451	145.7955	32.8%
2048	0.2148	78.2090	0.0992	169.4276	46.2%
4096	0.4705	142.7339	0.3785	177.4489	80.4%

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

See also

• sgemv.js — WebGPU benchmark script
• sgemv.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	0.0136	0.3302	0.0051	0.8833	37.4%
64	0.0107	1.6048	0.0051	3.3605	47.8%
128	0.0150	4.4834	0.0055	12.2216	36.7%
256	0.0219	12.1081	0.0103	25.6992	47.1%
512	0.0505	20.9004	0.0115	91.5556	22.8%
1024	0.1106	38.0370	0.0327	128.8153	29.5%
1280	0.1363	48.2048	0.0451	145.7955	33.1%
2048	0.2187	76.8243	0.0999	168.1255	45.7%
4096	0.4732	141.9231	0.3725	180.2687	78.7%

Nvidia Geforce Gtx 1650 — stride = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0137	0.3275	0.0051	0.8777	37.3%
64	0.0108	1.5811	0.0052	3.3189	47.6%
128	0.0149	4.4979	0.0058	11.6444	38.6%
256	0.0223	11.9166	0.0106	25.1152	47.4%
512	0.0576	18.3111	0.0148	71.3420	25.7%
1024	0.1096	38.3813	0.0334	125.7952	30.5%
1280	0.1364	48.1539	0.0468	140.5065	34.3%
2048	0.2148	78.2265	0.1020	164.7234	47.5%
4096	0.4778	140.5450	0.3826	175.5490	80.1%

Nvidia Geforce Gtx 1650 — stride = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0137	0.3260	0.0054	0.8284	39.3%
64	0.0113	1.5184	0.0053	3.2096	47.3%
128	0.0155	4.3216	0.0060	11.2688	38.4%
256	0.0221	11.9942	0.0104	25.5015	47.0%
512	0.0584	18.0652	0.0156	67.7492	26.7%
1024	0.1105	38.0811	0.0344	122.3985	31.1%
1280	0.1374	47.8062	0.0471	139.4565	34.3%
2048	0.2166	77.5849	0.1027	163.6708	47.4%
4096	0.4799	139.9452	0.3861	173.9556	80.4%

See also:

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

Transpose sweep

Unless noted otherwise, every result above uses trans = "no-transpose". trans = "transpose"'s parallelism is bounded by n (one workgroup per output-column tile) rather than m, so it's slower at matched square shapes and substantially slower on tall-narrow shapes — this section sweeps every (m, n) pair for both trans values to characterize that shape sensitivity, not just a single square-shape A/B. Collapsed by default since it's 18 shape combinations — expand a trans value, then a shape, to see its table and chart.

Nvidia Geforce Gtx 1650 — trans = no-transpose (9 shapes)

m = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5469	0.0055	0.8092	67.6%
64	0.0082	1.0625	0.0054	1.6000	66.4%
128	0.0084	2.0458	0.0056	3.0455	67.2%
256	0.0084	4.0611	0.0080	4.2390	95.8%
512	0.0089	7.6534	0.0084	8.0303	95.3%
1024	0.0095	14.2973	0.0085	15.9397	89.7%
1280	0.0100	16.8408	0.0087	19.4412	86.6%
2048	0.0103	26.2202	0.0101	26.7595	98.0%
4096	0.0123	44.0208	0.0105	51.6153	85.3%

m = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	1.0781	0.0053	1.6727	64.5%
64	0.0083	2.0695	0.0052	3.3189	62.4%
128	0.0084	4.0152	0.0059	5.7705	69.6%
256	0.0085	7.8797	0.0082	8.1875	96.2%
512	0.0092	14.5000	0.0087	15.3529	94.4%
1024	0.0101	26.4216	0.0093	28.5969	92.4%
1280	0.0102	32.5500	0.0098	34.1508	95.3%
2048	0.0110	48.2783	0.0112	47.5886	101.4%
4096	0.0164	65.0313	0.0139	76.7189	84.8%

m = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0086	2.0410	0.0053	3.2913	62.0%
64	0.0087	3.9262	0.0056	6.0455	64.9%
128	0.0092	7.3159	0.0059	11.4536	63.9%
256	0.0092	14.4194	0.0097	13.7068	105.2%
512	0.0101	26.3111	0.0107	24.7773	106.2%
1024	0.0107	49.3851	0.0119	44.3539	111.3%
1280	0.0116	57.1840	0.0123	53.9739	105.9%
2048	0.0148	71.6273	0.0162	65.1992	109.9%
4096	0.0215	98.1872	0.0222	95.2161	103.1%

m = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0102	3.4125	0.0099	3.5397	96.4%
64	0.0102	6.6250	0.0101	6.6877	99.1%
128	0.0107	12.4471	0.0106	12.6545	98.4%
256	0.0116	22.9584	0.0113	23.4124	98.1%
512	0.0123	43.0000	0.0129	40.8713	105.2%
1024	0.0164	64.5010	0.0173	61.0936	105.6%
1280	0.0192	68.4688	0.0199	66.2122	103.4%
2048	0.0246	85.7500	0.0254	82.9421	103.4%
4096	0.0373	113.0511	0.0406	103.8233	108.9%

m = 512

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0123	5.6771	0.0100	6.9872	81.2%
64	0.0124	10.9072	0.0103	13.2044	82.6%
128	0.0138	19.3635	0.0104	25.6888	75.4%
256	0.0143	36.9286	0.0123	43.0833	85.7%
512	0.0164	64.3750	0.0159	66.2513	97.2%
1024	0.0246	85.6667	0.0228	92.1457	93.0%
1280	0.0269	97.9250	0.0261	100.6834	97.3%
2048	0.0360	116.8498	0.0367	114.7586	101.8%
4096	0.0598	140.5262	0.0676	124.4832	112.9%

m = 1024

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0165	8.4419	0.0100	13.9839	60.4%
64	0.0167	16.1992	0.0104	26.0988	62.1%
128	0.0180	29.6898	0.0122	43.6592	68.0%
256	0.0197	53.6623	0.0148	71.3952	75.2%
512	0.0251	83.9465	0.0225	93.5455	89.7%
1024	0.0358	117.3714	0.0347	121.1576	96.9%
1280	0.0422	124.4364	0.0409	128.5258	96.8%
2048	0.0639	131.5253	0.0594	141.5172	92.9%
4096	0.1100	152.8102	0.1127	149.1425	102.5%

m = 1280

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0246	7.0793	0.0100	17.3930	40.7%
64	0.0263	12.8564	0.0107	31.6407	40.6%
128	0.0282	23.6545	0.0118	56.6422	41.8%
256	0.0307	43.0782	0.0175	75.4557	57.1%
512	0.0368	71.5065	0.0246	107.1667	66.7%
1024	0.0492	106.9235	0.0398	132.2237	80.9%
1280	0.0573	114.5536	0.0464	141.6212	80.9%
2048	0.0794	132.2546	0.0676	155.4242	85.1%
4096	0.1413	148.6279	0.1272	165.0797	90.0%

m = 2048

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0369	7.5557	0.0065	42.6863	17.7%
64	0.0370	14.6165	0.0078	69.5638	21.0%
128	0.0435	24.5004	0.0119	89.7466	27.3%
256	0.0451	46.9318	0.0206	102.4496	45.8%
512	0.0538	78.3386	0.0346	121.7272	64.4%
1024	0.0750	112.1092	0.0550	152.7814	73.4%
1280	0.0873	120.3195	0.0684	153.5791	78.3%
2048	0.1211	138.7202	0.1004	167.4019	82.9%
4096	0.2432	138.1143	0.2166	155.0713	89.1%

m = 4096

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0696	8.0055	0.0185	30.1507	26.6%
64	0.0737	14.6765	0.0246	44.0104	33.3%
128	0.0839	25.4010	0.0328	65.0156	39.1%
256	0.0842	50.2006	0.0483	87.4730	57.4%
512	0.1017	82.7904	0.0676	124.6364	66.4%
1024	0.1469	114.4500	0.1283	131.0488	87.3%
1280	0.1731	121.3463	0.1436	146.2725	83.0%
2048	0.2526	133.0026	0.1985	169.2361	78.6%
4096	0.4726	142.1057	0.3776	177.8775	79.9%

Nvidia Geforce Gtx 1650 — trans = transpose (9 shapes)

m = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0226	0.1984	0.0058	0.7735	25.7%
64	0.0225	0.3920	0.0100	0.8818	44.5%
128	0.0226	0.7773	0.0045	3.9004	19.9%
256	0.0236	1.4837	0.0052	6.6994	22.1%
512	0.0239	2.9144	0.0052	13.2927	21.9%
1024	0.0236	5.9145	0.0061	22.8661	25.9%
1280	0.0241	7.2297	0.0062	28.1344	25.7%
2048	0.0240	11.6107	0.0078	35.9093	32.3%
4096	0.0242	23.0623	0.0101	55.1013	41.9%

m = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0184	0.4722	0.0047	1.8630	25.3%
64	0.0187	0.9194	0.0051	3.3817	27.2%
128	0.0186	1.8313	0.0047	7.2381	25.3%
256	0.0186	3.6520	0.0050	13.5463	27.0%
512	0.0188	7.1973	0.0056	24.3218	29.6%
1024	0.0196	13.8170	0.0061	44.0417	31.4%
1280	0.0199	16.9631	0.0064	52.5771	32.3%
2048	0.0200	27.0897	0.0081	67.0794	40.4%
4096	0.0261	41.4979	0.0110	98.2558	42.2%

m = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0246	0.6970	0.0049	3.5033	19.9%
64	0.0249	1.3573	0.0050	6.7910	20.0%
128	0.0264	2.5421	0.0052	12.9383	19.6%
256	0.0265	5.0435	0.0057	23.4607	21.5%
512	0.0264	10.1042	0.0061	43.9894	23.0%
1024	0.0280	19.0572	0.0076	70.1305	27.2%
1280	0.0328	20.3281	0.0079	84.2753	24.1%
2048	0.0636	16.7569	0.0098	108.2797	15.5%
4096	0.1230	17.3195	0.0167	127.9078	13.5%

m = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0425	0.8018	0.0053	6.4485	12.4%
64	0.0407	1.6491	0.0054	12.3658	13.3%
128	0.0410	3.2475	0.0056	23.7714	13.7%
256	0.0410	6.4750	0.0060	44.5591	14.5%
512	0.0411	12.8948	0.0066	80.7024	16.0%
1024	0.0860	12.2976	0.0083	127.6293	9.6%
1280	0.1109	11.9244	0.0125	105.9282	11.3%
2048	0.1249	16.9241	0.0168	125.7469	13.5%
4096	0.2214	19.0950	0.0285	148.1256	12.9%

m = 512

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0639	1.0621	0.0054	12.5444	8.5%
64	0.0660	2.0233	0.0055	24.1387	8.4%
128	0.0670	3.9570	0.0059	44.8000	8.8%
256	0.0673	7.8516	0.0064	81.9454	9.6%
512	0.1444	7.3017	0.0082	128.7500	5.7%
1024	0.1987	10.6082	0.0184	114.2342	9.3%
1280	0.2003	13.1518	0.0211	125.0821	10.5%
2048	0.2273	18.5341	0.0299	140.7247	13.2%
4096	0.4219	19.9660	0.0525	160.4584	12.4%

m = 1024

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.1230	1.1008	0.0062	21.7026	5.1%
64	0.1264	2.1101	0.0067	39.7900	5.3%
128	0.1283	4.1277	0.0070	75.3713	5.5%
256	0.2703	3.9015	0.0084	125.8015	3.1%
512	0.3686	5.7111	0.0171	123.4372	4.6%
1024	0.3819	11.0139	0.0320	131.3247	8.4%
1280	0.3912	13.4398	0.0361	145.7746	9.2%
2048	0.4379	19.2051	0.0515	163.2706	11.8%
4096	0.8195	20.5170	0.0976	172.2189	11.9%

m = 1280

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.1522	1.1117	0.0076	22.4068	5.0%
64	0.1430	2.3302	0.0076	44.1356	5.3%
128	0.1446	4.5760	0.0080	82.3586	5.6%
256	0.4303	3.0630	0.0135	97.5924	3.1%
512	0.4404	5.9740	0.0205	128.4500	4.7%
1024	0.4632	11.3468	0.0347	151.3880	7.5%
1280	0.4689	14.0085	0.0424	155.1039	9.0%
2048	0.5310	19.7868	0.0819	128.2625	15.4%
4096	0.2291	91.7217	0.1200	175.0084	52.4%

m = 2048

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.1929	1.4028	0.0083	32.6486	4.3%
64	0.1966	2.7105	0.0104	51.3282	5.3%
128	0.2554	4.1421	0.0110	95.8145	4.3%
256	0.3047	6.9166	0.0204	103.0610	6.7%
512	0.3032	13.8754	0.0287	146.5507	9.5%
1024	0.3123	26.9170	0.0530	158.6566	17.0%
1280	0.3236	32.4620	0.0635	165.4516	19.6%
2048	0.3279	51.2425	0.1006	167.0025	30.7%
4096	0.3484	96.4409	0.1856	180.9940	53.3%

m = 4096

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.3707	1.4594	0.0089	60.6966	2.4%
64	0.5186	2.0544	0.0104	102.1350	2.0%
128	0.5489	3.8526	0.0190	111.0588	3.5%
256	0.5492	7.6705	0.0307	137.1333	5.6%
512	0.5340	15.7469	0.0516	162.9163	9.7%
1024	0.5591	30.0513	0.1017	165.2679	18.2%
1280	0.5646	37.1886	0.1210	173.4810	21.4%
2048	0.5732	58.5927	0.1949	172.3198	34.0%
4096	0.6058	110.8569	0.3691	181.9488	60.9%

See also:

• trans.sgemv.js — WebGPU trans-sweep benchmark script
• trans.sgemv.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 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	0.0082	0.5313	0.0048	0.9067	58.6%
64	0.0083	2.0426	0.0049	3.4286	59.6%
128	0.0088	7.5636	0.0053	12.5680	60.2%
256	0.0116	22.7438	0.0100	26.5466	85.7%
512	0.0164	64.2500	0.0135	77.9526	82.4%
1024	0.0356	117.8887	0.0328	128.3127	91.9%
1280	0.0492	133.5417	0.0451	145.6818	91.7%
2048	0.1030	163.0067	0.0995	168.7188	96.6%
4096	0.3742	179.4310	0.3780	177.6083	101.0%

Nvidia Geforce Gtx 1650 — pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5313	0.0047	0.9220	57.6%
64	0.0084	2.0230	0.0048	3.5200	57.5%
128	0.0091	7.2982	0.0053	12.4551	58.6%
256	0.0114	23.1261	0.0102	25.9623	89.1%
512	0.0164	64.2500	0.0143	73.5106	87.4%
1024	0.0358	117.2571	0.0328	128.3127	91.4%
1280	0.0492	133.5417	0.0452	145.1150	92.0%
2048	0.1050	159.9512	0.1000	167.9898	95.2%
4096	0.3766	178.2950	0.3831	175.2496	101.7%

Nvidia Geforce Gtx 1650 — pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5313	0.0047	0.9347	56.8%
64	0.0082	2.0625	0.0049	3.4623	59.6%
128	0.0091	7.2982	0.0055	12.0930	60.4%
256	0.0116	22.7125	0.0101	26.0442	87.2%
512	0.0164	64.2500	0.0137	77.0398	83.4%
1024	0.0359	117.1525	0.0328	128.2500	91.3%
1280	0.0492	133.3680	0.0451	145.6818	91.5%
2048	0.1049	160.0488	0.1003	167.4003	95.6%
4096	0.3754	178.8345	0.3809	176.2506	101.5%

Nvidia Geforce Gtx 1650 — pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5313	0.0049	0.8947	59.4%
64	0.0082	2.0625	0.0049	3.4286	60.2%
128	0.0090	7.3628	0.0055	12.1637	60.5%
256	0.0113	23.3881	0.0102	25.8000	90.7%
512	0.0165	63.8757	0.0141	74.5941	85.6%
1024	0.0356	117.9946	0.0329	127.8754	92.3%
1280	0.0492	133.5417	0.0451	145.6818	91.7%
2048	0.1044	160.7843	0.1000	167.9629	95.7%
4096	0.3748	179.1552	0.4318	155.5011	115.2%

Nvidia Geforce Gtx 1650 — pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5313	0.0156	0.2784	190.8%
64	0.0083	2.0426	0.0157	1.0776	189.5%
128	0.0094	7.0508	0.0167	3.9923	176.6%
256	0.0114	23.0937	0.0183	14.4588	159.7%
512	0.0164	64.2500	0.0246	42.8333	150.0%
1024	0.0360	116.8919	0.0452	92.9756	125.7%
1280	0.0502	130.8163	0.0432	151.8845	86.1%
2048	0.1041	161.3776	0.1000	167.8823	96.1%
4096	0.3871	173.4531	0.3820	175.7487	98.7%

Nvidia Geforce Gtx 1650 — pad = 48

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0117	0.3721	0.0041	1.0625	35.0%
64	0.0117	1.4387	0.0048	3.4967	41.1%
128	0.0126	5.2926	0.0055	12.1637	43.5%
256	0.0164	16.1250	0.0102	25.8000	62.5%
512	0.0225	46.6941	0.0143	73.4286	63.6%
1024	0.0451	93.2727	0.0328	128.2500	72.7%
1280	0.0602	109.0774	0.0451	145.6301	74.9%
2048	0.1227	136.8984	0.1000	167.9898	81.5%
4096	0.4616	145.4441	0.3824	175.5796	82.8%

Nvidia Geforce Gtx 1650 — pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0123	0.3542	0.0049	0.8918	39.7%
64	0.0127	1.3266	0.0049	3.4623	38.3%
128	0.0142	4.6847	0.0054	12.3077	38.1%
256	0.0180	14.6383	0.0102	25.8000	56.7%
512	0.0262	40.2152	0.0142	74.2573	54.2%
1024	0.0497	84.4825	0.0328	128.2500	65.9%
1280	0.0675	97.1902	0.0451	145.6818	66.7%
2048	0.1344	124.9673	0.0996	168.6646	74.1%
4096	0.4628	145.0619	0.3809	176.2581	82.3%

Nvidia Geforce Gtx 1650 — pad = 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0126	0.3443	0.0048	0.9067	38.0%
64	0.0140	1.2096	0.0047	3.6289	33.3%
128	0.0147	4.5316	0.0054	12.3810	36.6%
256	0.0186	14.1978	0.0098	26.8488	52.9%
512	0.0280	37.5954	0.0136	77.2207	48.7%
1024	0.0533	78.8757	0.0328	128.2500	61.5%
1280	0.0715	91.7558	0.0451	145.4752	63.1%
2048	0.1415	118.6659	0.0982	171.0560	69.4%
4096	0.5120	131.1401	0.3827	175.4548	74.7%

See also:

• lda.sgemv.js — WebGPU lda-sweep benchmark script
• lda.sgemv.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	0.0082	0.5313	0.0036	1.2035	44.1%
64	0.0083	2.0386	0.0038	4.4936	45.4%
128	0.0091	7.2982	0.0041	16.2500	44.9%
256	0.0113	23.3881	0.0062	42.5567	55.0%
512	0.0164	64.2500	0.0123	85.6667	75.0%
1024	0.0360	116.6841	0.0307	136.8000	85.3%
1280	0.0491	133.6287	0.0451	145.6818	91.7%
2048	0.1035	162.3009	0.0994	168.9361	96.1%
4096	0.3742	179.4463	0.3800	176.6959	101.6%

Nvidia Geforce Gtx 1650 — alpha = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0067	0.6538	0.0021	2.1085	31.0%
64	0.0069	2.4444	0.0020	8.2500	29.6%
128	0.0085	7.8049	0.0020	32.5000	24.0%
256	0.0109	24.1757	0.0021	128.0000	18.9%
512	0.0159	66.2558	0.0020	514.0001	12.9%
1024	0.0354	118.5806	0.0026	1631.4036	7.3%
1280	0.0488	134.6373	0.0027	2441.9048	5.5%
2048	0.1026	163.6165	0.0027	6322.8916	2.6%
4096	0.3741	179.4847	0.0041	16456.2793	1.1%

Nvidia Geforce Gtx 1650 — alpha = 1e-38

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0107	0.4054	0.0050	0.8774	46.2%
64	0.0113	1.5000	0.0048	3.5556	42.2%
128	0.0123	5.4167	0.0056	11.8857	45.6%
256	0.0166	15.9382	0.0102	25.8000	61.8%
512	0.0226	46.6279	0.0141	74.6788	62.4%
1024	0.0451	93.2727	0.0328	128.2500	72.7%
1280	0.0583	112.6105	0.0451	145.4752	77.4%
2048	0.1167	143.8596	0.1000	167.9360	85.7%
4096	0.4588	146.3469	0.3781	177.5708	82.4%

Nvidia Geforce Gtx 1650 — alpha = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5313	0.0064	0.6834	77.7%
64	0.0083	2.0269	0.0064	2.6400	76.8%
128	0.0090	7.4021	0.0068	9.8345	75.3%
256	0.0114	23.1261	0.0127	20.8222	111.1%
512	0.0164	64.2500	0.0164	64.2500	100.0%
1024	0.0360	116.7360	0.0409	102.6401	113.7%
1280	0.0490	133.8903	0.0430	152.6190	87.7%
2048	0.1037	161.8754	0.0999	168.1782	96.3%
4096	0.3744	179.3466	0.3782	177.5482	101.0%

Nvidia Geforce Gtx 1650 — alpha = 2.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5313	0.0052	0.8421	63.1%
64	0.0083	2.0308	0.0047	3.5676	56.9%
128	0.0090	7.3890	0.0055	12.0580	61.3%
256	0.0113	23.2891	0.0100	26.3770	88.3%
512	0.0164	64.2500	0.0140	74.9339	85.7%
1024	0.0360	116.8919	0.0331	127.0097	92.0%
1280	0.0492	133.5417	0.0576	113.9556	117.2%
2048	0.1027	163.5655	0.0976	172.0092	95.1%
4096	0.3746	179.2547	0.3793	177.0164	101.3%

See also:

• alpha.sgemv.js — WebGPU alpha-sweep benchmark script
• alpha.sgemv.c — CUDA / cuBLAS alpha-sweep reference script

beta sweep

beta scales the existing y/C before accumulation. Reference BLAS is permitted to skip reading that operand entirely when beta is 0, so unlike alpha this sweep has a mechanism to be non-flat — a step at 0 means the shortcut is taken, and its size is what it saves.

Nvidia Geforce Gtx 1650 — beta = -3.75

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5313	0.0055	0.7953	66.8%
64	0.0083	2.0426	0.0050	3.3738	60.5%
128	0.0089	7.4552	0.0057	11.7183	63.6%
256	0.0114	23.1261	0.0104	25.3641	91.2%
512	0.0164	64.2500	0.0142	74.2573	86.5%
1024	0.0362	116.1167	0.0328	128.2500	90.5%
1280	0.0490	133.9341	0.0451	145.6818	91.9%
2048	0.1029	163.2602	0.0998	168.2860	97.0%
4096	0.3741	179.4770	0.4267	157.3553	114.1%

Nvidia Geforce Gtx 1650 — beta = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5313	0.0048	0.9097	58.4%
64	0.0083	2.0386	0.0047	3.5676	57.1%
128	0.0090	7.4286	0.0056	11.8857	62.5%
256	0.0114	23.0937	0.0100	26.3770	87.6%
512	0.0164	64.2500	0.0136	77.4024	83.0%
1024	0.0360	116.6841	0.0328	128.2500	91.0%
1280	0.0491	133.6287	0.0448	146.4620	91.2%
2048	0.1032	162.7035	0.0991	169.3996	96.0%
4096	0.3742	179.4233	0.3785	177.3681	101.2%

Nvidia Geforce Gtx 1650 — beta = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5313	0.0052	0.8293	64.1%
64	0.0083	2.0465	0.0052	3.2294	63.4%
128	0.0090	7.3628	0.0057	11.6201	63.4%
256	0.0115	22.9333	0.0102	25.8809	88.6%
512	0.0164	64.2500	0.0141	74.5096	86.2%
1024	0.0361	116.3739	0.0328	128.2500	90.7%
1280	0.0491	133.7594	0.0453	144.8076	92.4%
2048	0.1033	162.5019	0.0997	168.3670	96.5%
4096	0.3741	179.4770	0.3806	176.4137	101.7%

Nvidia Geforce Gtx 1650 — beta = 2.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0082	0.5313	0.0053	0.8168	65.0%
64	0.0083	2.0386	0.0049	3.4286	59.5%
128	0.0089	7.4419	0.0056	11.8182	63.0%
256	0.0112	23.5886	0.0101	26.2095	90.0%
512	0.0164	64.2500	0.0237	44.3641	144.8%
1024	0.0362	116.2195	0.0442	95.0963	122.2%
1280	0.0490	134.0216	0.0572	114.7524	116.8%
2048	0.1036	162.1505	0.1001	167.6945	96.7%
4096	0.3744	179.3543	0.3805	176.4433	101.6%

See also:

• beta.sgemv.js — WebGPU beta-sweep benchmark script
• beta.sgemv.c — CUDA / cuBLAS beta-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	0.0136	0.3200
64	0.0105	1.6024
128	0.0150	4.4492
256	0.0216	12.2584
512	0.0503	20.9462
1024	0.1086	38.6942
1280	0.1372	47.8358
2048	0.2150	78.0952
4096	0.4722	142.2010

Nvidia Geforce Gtx 1650 — layout = row-major

n	compute ms	GB/s
32	0.0082	0.5313
64	0.0084	2.0153
128	0.0089	7.5090
256	0.0114	23.2563
512	0.0164	64.3128
1024	0.0360	116.5806
1280	0.0488	134.5049
2048	0.1036	162.1755
4096	0.3738	179.6307

See also:

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