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

Benchmark results for sgemm 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	0.0088	1.8720	0.0076	2.1603	86.7%
64	0.0111	5.9020	0.0082	8.0000	73.8%
128	0.0186	14.0635	0.0104	25.2840	55.6%
256	0.0696	15.0588	0.0348	30.1038	50.0%
512	0.1714	24.4674	0.1544	27.1596	90.1%
1024	1.1325	14.8140	0.7723	21.7245	68.2%

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

See also

• sgemm.js — WebGPU benchmark script
• sgemm.c — CUDA / cuBLAS reference script

Transpose sweep

Unless noted otherwise, every result above uses transA = transB = "no-transpose". Both shaders load A/B into shared memory with a transpose-dependent index that scatters what would otherwise be a coalesced load — but it's asymmetric: transB dominates (measured +22-57% at n=1024) while transA is small and can even be faster than no-transpose, because B's tile dimension spans a full warp in the coalesced case (so transpose scatters every warp) while A's never gets a full-warp-coalesced load to begin with. All 4 (transA, transB) combinations are swept — collapsed below by default, expand a transA value, then a transB, to see its table and chart (4 combinations total).

Nvidia Geforce Gtx 1650 — transA = no-transpose (2 transB values)

transB = no-transpose

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0088	1.8652	0.0077	2.1378	87.2%
64	0.0114	5.7690	0.0082	7.9844	72.3%
128	0.0185	14.1608	0.0106	24.7492	57.2%
256	0.0696	15.0588	0.0355	29.5473	51.0%
512	0.1720	24.3810	0.1551	27.0447	90.2%
1024	1.1313	14.8297	0.7757	21.6295	68.6%

transB = transpose

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0090	1.8221	0.0084	1.9431	93.8%
64	0.0117	5.6110	0.0087	7.5294	74.5%
128	0.0188	13.9676	0.0108	24.3086	57.5%
256	0.0712	14.7272	0.0289	36.2879	40.6%
512	0.1902	22.0567	0.1546	27.1230	81.3%
1024	1.2870	13.0358	0.7701	21.7850	59.8%

Nvidia Geforce Gtx 1650 — transA = transpose (2 transB values)

transB = no-transpose

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0099	1.6490	0.0061	2.6667	61.8%
64	0.0123	5.3333	0.0070	9.3730	56.9%
128	0.0205	12.8000	0.0108	24.3447	52.6%
256	0.0737	14.2222	0.0424	24.7399	57.5%
512	0.2097	20.0034	0.1978	21.2005	94.4%
1024	1.3873	12.0935	0.7915	21.1958	57.1%

transB = transpose

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0099	1.6623	0.0063	2.6189	63.5%
64	0.0125	5.2379	0.0077	8.5511	61.3%
128	0.0206	12.7008	0.0096	27.3980	46.4%
256	0.0752	13.9527	0.0466	22.4823	62.1%
512	0.2293	18.2921	0.1847	22.7043	80.6%
1024	1.5372	10.9139	0.7722	21.7276	50.2%

See also:

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

Ldb sweep

Unless noted otherwise, every result above uses a tight lda/ldb/ldc (no padding). lda and ldc were scoped and found to be non-effects; padding ldb only matters for transB = "transpose" here (swept at both transB values below so that's visible in the data). Collapsed below by default — expand a transB value, then a pad, to see its table and chart.

Nvidia Geforce Gtx 1650 — transB = no-transpose (6 pads)

pad = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0088	1.8686	0.0077	2.1333	87.6%
64	0.0116	5.6497	0.0082	8.0000	70.6%
128	0.0186	14.1120	0.0104	25.2450	55.9%
256	0.0696	15.0588	0.0349	30.0073	50.2%
512	0.1707	24.5660	0.1538	27.2783	90.1%
1024	1.1319	14.8219	0.7722	21.7276	68.2%

pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0095	1.7239	0.0073	2.2456	76.8%
64	0.0119	5.4906	0.0082	8.0000	68.6%
128	0.0189	13.8847	0.0112	23.4057	59.3%
256	0.0655	16.0000	0.0371	28.2605	56.6%
512	0.2021	20.7540	0.1643	25.5327	81.3%
1024	1.2963	12.9422	0.7968	21.0545	61.5%

pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0088	1.8652	0.0067	2.4498	76.1%
64	0.0116	5.6264	0.0077	8.4979	66.2%
128	0.0186	14.1120	0.0101	26.0477	54.2%
256	0.0698	15.0209	0.0343	30.5387	49.2%
512	0.1721	24.3742	0.1488	28.1784	86.5%
1024	1.1434	14.6728	0.7691	21.8149	67.3%

pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0087	1.8789	0.0061	2.6667	70.5%
64	0.0110	5.9622	0.0079	8.2915	71.9%
128	0.0185	14.1608	0.0102	25.6000	55.3%
256	0.0699	14.9968	0.0370	28.3338	52.9%
512	0.1720	24.3810	0.1591	26.3594	92.5%
1024	1.1376	14.7477	0.7731	21.7011	68.0%

pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0089	1.8417	0.0065	2.5347	72.7%
64	0.0115	5.7047	0.0082	8.0000	71.3%
128	0.0185	14.1485	0.0128	20.4035	69.3%
256	0.0696	15.0588	0.0364	28.8324	52.2%
512	0.1718	24.4082	0.1647	25.4707	95.8%
1024	1.1273	14.8823	0.7815	21.4683	69.3%

pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0088	1.8517	0.0079	2.0813	89.0%
64	0.0114	5.7367	0.0082	8.0000	71.7%
128	0.0184	14.2222	0.0104	25.3230	56.2%
256	0.0696	15.0588	0.0351	29.8842	50.4%
512	0.1716	24.4355	0.1562	26.8508	91.0%
1024	1.1278	14.8766	0.7747	21.6558	68.7%

Nvidia Geforce Gtx 1650 — transB = transpose (6 pads)

pad = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0089	1.8484	0.0090	1.8221	101.4%
64	0.0116	5.6575	0.0100	6.5746	86.1%
128	0.0191	13.7334	0.0102	25.6401	53.6%
256	0.0706	14.8574	0.0287	36.5714	40.6%
512	0.1902	22.0567	0.1596	26.2748	83.9%
1024	1.2882	13.0237	0.7696	21.7995	59.7%

pad = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0089	1.8384	0.0068	2.4208	75.9%
64	0.0109	6.0147	0.0075	8.7897	68.4%
128	0.0195	13.4185	0.0100	26.3408	50.9%
256	0.0716	14.6351	0.0269	38.9863	37.5%
512	0.2110	19.8790	0.1447	28.9918	68.6%
1024	1.5299	10.9665	0.7639	21.9625	49.9%

pad = 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0087	1.8824	0.0070	2.3486	80.1%
64	0.0120	5.4613	0.0078	8.3934	65.1%
128	0.0197	13.3203	0.0116	22.5986	58.9%
256	0.0696	15.0692	0.0285	36.7973	41.0%
512	0.1841	22.7793	0.1518	27.6261	82.5%
1024	1.2446	13.4801	0.7700	21.7872	61.9%

pad = 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0089	1.8484	0.0068	2.4265	76.2%
64	0.0114	5.7367	0.0078	8.4280	68.1%
128	0.0188	13.9201	0.0122	21.5013	64.7%
256	0.0698	15.0277	0.0270	38.8017	38.7%
512	0.1864	22.5055	0.1514	27.7020	81.2%
1024	1.2702	13.2081	0.7699	21.7913	60.6%

pad = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0088	1.8517	0.0096	1.7038	108.7%
64	0.0118	5.5652	0.0082	7.9534	70.0%
128	0.0192	13.6420	0.0121	21.5863	63.2%
256	0.0706	14.8439	0.0265	39.5271	37.6%
512	0.2058	20.3844	0.1468	28.5778	71.3%
1024	1.4122	11.8799	0.7639	21.9625	54.1%

pad = 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0095	1.7181	0.0089	1.8450	93.1%
64	0.0123	5.3195	0.0083	7.8618	67.7%
128	0.0205	12.7601	0.0123	21.3890	59.7%
256	0.0763	13.7421	0.0287	36.5714	37.6%
512	0.2063	20.3275	0.1631	25.7130	79.1%
1024	1.4701	11.4126	0.7682	21.8390	52.3%

See also:

• ldb.sgemm.js — WebGPU ldb-sweep benchmark script
• ldb.sgemm.c — CUDA / cuBLAS ldb-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
64	0.0123	5.3473	0.0079	8.3083	64.4%
128	0.0195	13.4516	0.0116	22.6925	59.3%
256	0.0704	14.8878	0.0399	26.2775	56.7%
512	0.1727	24.2906	0.1659	25.2840	96.1%
1024	1.1614	14.4456	0.7810	21.4811	67.2%
2048	9.7422	6.8885	7.7346	8.6765	79.4%

Nvidia Geforce Gtx 1650 — alpha = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0113	5.7935	0.0040	16.4498	35.2%
128	0.0185	14.1853	0.0032	81.5124	17.4%
256	0.0696	15.0588	0.0037	284.9391	5.3%
512	0.1717	24.4264	0.0082	512.0000	4.8%
1024	1.1357	14.7724	0.0249	673.8920	2.2%
2048	9.6645	6.9439	0.0924	726.4121	1.0%

Nvidia Geforce Gtx 1650 — alpha = 1e-38

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0126	5.1848	0.0076	8.5870	60.4%
128	0.0197	13.3095	0.0103	25.5203	52.2%
256	0.0704	14.8945	0.0363	28.9214	51.5%
512	0.1725	24.3154	0.1585	26.4551	91.9%
1024	1.1757	14.2694	0.8049	20.8448	68.5%
2048	9.6668	6.9422	7.3627	9.1147	76.2%

Nvidia Geforce Gtx 1650 — alpha = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0127	5.1457	0.0082	8.0000	64.3%
128	0.0200	13.1177	0.0106	24.7492	53.0%
256	0.0706	14.8540	0.0339	30.9424	48.0%
512	0.1727	24.2928	0.1516	27.6757	87.8%
1024	1.1674	14.3717	0.7844	21.3882	67.2%
2048	9.7573	6.8778	7.3914	9.0794	75.8%

Nvidia Geforce Gtx 1650 — alpha = 2.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0123	5.3333	0.0091	7.2240	73.8%
128	0.0198	13.2342	0.0119	22.0809	59.9%
256	0.0704	14.9047	0.0397	26.4365	56.4%
512	0.1723	24.3402	0.1667	25.1626	96.7%
1024	1.1680	14.3644	0.8499	19.7394	72.8%
2048	9.7554	6.8791	7.4711	8.9825	76.6%

See also:

• alpha.sgemm.js — WebGPU alpha-sweep benchmark script
• alpha.sgemm.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
64	0.0126	5.2046	0.0082	8.0000	65.1%
128	0.0199	13.2023	0.0120	21.8453	60.4%
256	0.0707	14.8372	0.0418	25.0904	59.1%
512	0.1722	24.3560	0.1676	25.0233	97.3%
1024	1.1762	14.2640	0.7881	21.2892	67.0%
2048	9.7557	6.8790	7.3085	9.1823	74.9%

Nvidia Geforce Gtx 1650 — beta = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0114	5.7609	0.0082	7.9689	72.3%
128	0.0187	14.0394	0.0105	24.9756	56.2%
256	0.0696	15.0588	0.0349	30.0348	50.1%
512	0.1705	24.6006	0.1542	27.1934	90.5%
1024	1.1315	14.8271	0.7908	21.2151	69.9%
2048	9.6724	6.9382	7.5027	8.9446	77.6%

Nvidia Geforce Gtx 1650 — beta = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0128	5.1328	0.0092	7.0988	72.3%
128	0.0212	12.3840	0.0122	21.5296	57.5%
256	0.0700	14.9694	0.0430	24.3810	61.4%
512	0.1722	24.3538	0.1700	24.6747	98.7%
1024	1.1835	14.1765	0.8043	20.8593	68.0%
2048	9.7527	6.8811	7.5445	8.8951	77.4%

Nvidia Geforce Gtx 1650 — beta = 2.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0123	5.3333	0.0094	6.9779	76.4%
128	0.0191	13.7219	0.0123	21.3890	64.2%
256	0.0715	14.6580	0.0436	24.0323	61.0%
512	0.1721	24.3674	0.1693	24.7773	98.3%
1024	1.1765	14.2601	0.8493	19.7546	72.2%
2048	9.6698	6.9401	7.2970	9.1968	75.5%

See also:

• beta.sgemm.js — WebGPU beta-sweep benchmark script
• beta.sgemm.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
64	0.0123	5.3333
128	0.0194	13.5070
256	0.0716	14.6515
512	0.1724	24.3244
1024	1.1837	14.1730
2048	9.6700	6.9399

Nvidia Geforce Gtx 1650 — layout = row-major

n	compute ms	GB/s
64	0.0115	5.7207
128	0.0188	13.9795
256	0.0697	15.0450
512	0.1712	24.4994
1024	1.1501	14.5871
2048	9.6660	6.9428

See also:

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

ldc sweep

Padding on the output matrix. C is written rather than streamed, so this measures write coalescing rather than read bandwidth — the row byte-stride is ldc*4, and a pad that moves it off the 128-byte boundary is what would show up here.

Nvidia Geforce Gtx 1650 — ldc = n + 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0136	4.8188	0.0080	8.2084	58.7%
128	0.0243	10.8003	0.0104	25.2062	42.8%
256	0.0889	11.7998	0.0348	30.1176	39.2%
512	0.2172	19.3136	0.1540	27.2273	70.9%
1024	1.4515	11.5583	0.7756	21.6304	53.4%
2048	9.7579	6.8774	6.8297	9.8260	70.0%

Nvidia Geforce Gtx 1650 — ldc = n + 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0123	5.3333	0.0089	7.3802	72.3%
128	0.0193	13.5629	0.0105	24.8619	54.6%
256	0.0709	14.7837	0.0377	27.7813	53.2%
512	0.1755	23.9008	0.1644	25.5128	93.7%
1024	1.1612	14.4480	0.7815	21.4683	67.3%
2048	9.7606	6.8755	9.4662	7.0893	97.0%

Nvidia Geforce Gtx 1650 — ldc = n + 8

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0125	5.2379	0.0088	7.4473	70.3%
128	0.0200	13.1282	0.0107	24.5269	53.5%
256	0.0705	14.8709	0.0370	28.3583	52.4%
512	0.1742	24.0808	0.1618	25.9215	92.9%
1024	1.1768	14.2566	0.7808	21.4877	66.3%
2048	9.7240	6.9014	9.0129	7.4459	92.7%

Nvidia Geforce Gtx 1650 — ldc = n + 16

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0123	5.3333	0.0082	8.0000	66.7%
128	0.0192	13.6306	0.0108	24.1652	56.4%
256	0.0708	14.8070	0.0393	26.6623	55.5%
512	0.1745	24.0389	0.1601	26.1961	91.8%
1024	1.1880	14.1226	0.8498	19.7435	71.5%
2048	9.7605	6.8755	8.9928	7.4625	92.1%

Nvidia Geforce Gtx 1650 — ldc = n + 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0127	5.1457	0.0078	8.4280	61.1%
128	0.0189	13.8495	0.0105	24.9756	55.5%
256	0.0704	14.8878	0.0387	27.0810	55.0%
512	0.1734	24.1875	0.1639	25.5950	94.5%
1024	1.1777	14.2458	0.8497	19.7438	72.2%
2048	9.7561	6.8787	9.0177	7.4419	92.4%

Nvidia Geforce Gtx 1650 — ldc = n + 48

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0124	5.2716	0.0076	8.6232	61.1%
128	0.0196	13.3638	0.0105	25.0520	53.3%
256	0.0704	14.8844	0.0386	27.1821	54.8%
512	0.1751	23.9576	0.1633	25.6903	93.3%
1024	1.1719	14.3164	0.8336	20.1258	71.1%
2048	9.7616	6.8748	9.5036	7.0614	97.4%

Nvidia Geforce Gtx 1650 — ldc = n + 64

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0124	5.2988	0.0076	8.5870	61.7%
128	0.0190	13.7796	0.0107	24.4173	56.4%
256	0.0703	14.9149	0.0401	26.1308	57.1%
512	0.1733	24.1964	0.1669	25.1264	96.3%
1024	1.1840	14.1701	0.7840	21.3991	66.2%
2048	9.6779	6.9342	9.0132	7.4457	93.1%

Nvidia Geforce Gtx 1650 — ldc = n + 128

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0123	5.3333	0.0073	8.9628	59.5%
128	0.0192	13.6193	0.0109	24.0587	56.6%
256	0.0699	15.0002	0.0402	26.0996	57.5%
512	0.1736	24.1563	0.1668	25.1433	96.1%
1024	1.1763	14.2625	0.7854	21.3620	66.8%
2048	9.7572	6.8779	9.3659	7.1652	96.0%

See also:

• ldc.sgemm.js — WebGPU ldc-sweep benchmark script
• ldc.sgemm.c — CUDA / cuBLAS ldc-sweep reference script