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

Benchmark results for drot 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	0.0069	0.1475	0.0035	0.2963	49.8%
64	0.0068	0.3019	0.0040	0.5161	58.5%
128	0.0067	0.6139	0.0037	1.1179	54.9%
512	0.0068	2.4151	0.0041	4.0315	59.9%
1024	0.0068	4.8075	0.0041	8.0000	60.1%
4096	0.0298	4.4019	0.0041	31.7519	13.9%
16384	0.0119	44.2213	0.0062	84.4536	52.4%
65536	0.0181	115.6858	0.0141	148.9454	77.7%
262144	0.0567	148.0203	0.0524	159.9414	92.5%
1048576	0.2031	165.1951	0.1965	170.7222	96.8%
4194304	0.7900	169.9027	0.7741	173.3757	98.0%
16777216	3.1922	168.1800	3.1308	171.4778	98.1%

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

See also

• drot.js — WebGPU benchmark script
• drot.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	0.0087	0.1172	0.0031	0.3333	35.2%
64	0.0074	0.2753	0.0031	0.6667	41.3%
128	0.0072	0.5651	0.0032	1.2736	44.4%
512	0.0069	2.3594	0.0036	4.4912	52.5%
1024	0.0072	4.5714	0.0038	8.6050	53.1%
4096	0.0087	15.0312	0.0042	31.5077	47.7%
16384	0.0192	27.3294	0.0086	61.2486	44.6%
65536	0.0700	29.9799	0.0553	37.9479	79.0%
262144	0.2874	29.1855	0.2110	39.7489	73.4%
1048576	1.1606	28.9107	0.8362	40.1254	72.1%
4194304	4.6374	28.9425	3.4979	38.3711	75.4%

Nvidia Geforce Gtx 1650 — stride = 5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0071	0.1445	0.0034	0.3019	47.9%
64	0.0068	0.3033	0.0032	0.6432	47.2%
128	0.0069	0.5926	0.0033	1.2427	47.7%
512	0.0071	2.3063	0.0038	4.3390	53.2%
1024	0.0075	4.3854	0.0038	8.6050	51.0%
4096	0.0092	14.1730	0.0041	31.7519	44.6%
16384	0.0215	24.3628	0.0182	28.8705	84.4%
65536	0.0937	22.3749	0.0683	30.7032	72.9%
262144	0.3859	21.7394	0.2683	31.2672	69.5%
1048576	1.5442	21.7294	1.0658	31.4831	69.0%

Nvidia Geforce Gtx 1650 — stride = 32

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
32	0.0075	0.1359	0.0031	0.3299	41.2%
64	0.0072	0.2838	0.0034	0.5953	47.7%
128	0.0076	0.5367	0.0038	1.0802	49.7%
512	0.0321	0.5102	0.0044	3.7372	13.7%
1024	0.0112	2.9174	0.0042	7.7576	37.6%
4096	0.0201	6.5068	0.0063	20.6869	31.5%
16384	0.0710	7.3835	0.0411	12.7552	57.9%
65536	0.2774	7.5607	0.1826	11.4834	65.8%
262144	1.0891	7.7026	0.7503	11.1810	68.9%

Nvidia Geforce Gtx 1650 — stride = 33

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
64	0.0073	0.2813	0.0032	0.6497	43.3%
128	0.0075	0.5482	0.0034	1.2133	45.2%
512	0.0080	2.0562	0.0041	4.0000	51.4%
1024	0.0082	4.0000	0.0041	8.0000	50.0%
4096	0.0205	6.4000	0.0098	13.3420	48.0%
16384	0.0756	6.9394	0.0475	11.0404	62.9%
65536	0.3072	6.8267	0.2132	9.8343	69.4%
262144	1.2512	6.7047	0.8825	9.5052	70.5%

Nvidia Geforce Gtx 1650 — stride = 255

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.0349	0.4699	0.0041	4.0000	11.7%
1024	0.0122	2.6771	0.0045	7.2367	37.0%
4096	0.0307	4.2667	0.0124	10.6114	40.2%
16384	0.1276	4.1083	0.0718	7.3029	56.3%
65536	0.5898	3.5556	0.3020	6.9453	51.2%

Nvidia Geforce Gtx 1650 — stride = 256

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
512	0.0080	2.0398	0.0042	3.9385	51.8%
1024	0.0082	4.0000	0.0042	7.8769	50.8%
4096	0.0278	4.7135	0.0099	13.2557	35.6%
16384	0.1124	4.6651	0.0696	7.5294	62.0%
65536	0.5636	3.7209	0.2930	7.1585	52.0%

See also:

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

c sweep

The cosine half of the plane rotation, swept with s held fixed so the two halves are attributed separately. srot's kernel computes both outputs unconditionally, so a flat sweep is expected; a step at c = 0 or c = 1 would mean an identity case is being short-circuited, which BLAS does not promise.

Nvidia Geforce Gtx 1650 — c = -0.75

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
1024	0.0069	4.7739	0.0036	9.2252	51.7%
65536	0.0177	118.7246	0.0110	189.9594	62.5%
1048576	0.2030	165.2602	0.1987	168.9072	97.8%
16777216	3.1680	169.4685	3.1366	171.1638	99.0%

Nvidia Geforce Gtx 1650 — c = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
1024	0.0070	4.6865	0.0035	9.3091	50.3%
65536	0.0176	118.9401	0.0111	189.6845	62.7%
1048576	0.2030	165.2732	0.1986	168.9344	97.8%
16777216	3.1802	168.8154	3.1314	171.4480	98.5%

Nvidia Geforce Gtx 1650 — c = 0.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
1024	0.0069	4.7628	0.0036	9.2252	51.6%
65536	0.0178	117.8705	0.0112	188.0516	62.7%
1048576	0.2029	165.3384	0.1986	168.9889	97.8%
16777216	3.1479	170.5470	3.1334	171.3360	99.5%

Nvidia Geforce Gtx 1650 — c = 0.7071067811865476

n	compute ms	GB/s
1024	0.0074	4.4043
65536	0.0177	118.6172
1048576	0.2036	164.7798
16777216	3.1840	168.6169

Nvidia Geforce Gtx 1650 — c = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
1024	0.0082	4.0000	0.0036	9.2252	43.4%
65536	0.0174	120.3600	0.0110	189.9594	63.4%
1048576	0.2031	165.2211	0.1987	168.9072	97.8%
16777216	3.1705	169.3343	3.1331	171.3552	98.8%

See also:

• cosine.drot.js — WebGPU c-sweep benchmark script
• cosine.drot.c — CUDA / cuBLAS c-sweep reference script

s sweep

The sine half of the plane rotation, swept with c held fixed — the counterpart to the cosine sweep. s = 0 makes the rotation an identity in exact arithmetic but is still fully computed and written, so a step there would indicate a short-circuit rather than a property of the maths.

Nvidia Geforce Gtx 1650 — s = -0.75

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
1024	0.0069	4.7739	0.0037	8.8276	54.1%
65536	0.0177	118.5099	0.0114	184.6084	64.2%
1048576	0.2030	165.2732	0.1987	168.9072	97.8%
16777216	3.1501	170.4283	3.1394	171.0102	99.7%

Nvidia Geforce Gtx 1650 — s = 0

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
1024	0.0067	4.8646	0.0041	8.0000	60.8%
65536	0.0175	119.9195	0.0113	185.9177	64.5%
1048576	0.2030	165.2993	0.1987	168.9072	97.9%
16777216	3.1520	170.3262	4.3451	123.5590	137.9%

Nvidia Geforce Gtx 1650 — s = 0.5

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
1024	0.0069	4.7189	0.0039	8.4628	55.8%
65536	0.0174	120.2495	0.0115	183.0614	65.7%
1048576	0.2031	165.2472	0.1986	168.9344	97.8%
16777216	3.1522	170.3167	3.3054	162.4204	104.9%

Nvidia Geforce Gtx 1650 — s = 0.7071067811865475

n	compute ms	GB/s
1024	0.0068	4.8075
65536	0.0178	118.0829
1048576	0.2031	165.2342
16777216	3.1529	170.2778

Nvidia Geforce Gtx 1650 — s = 1

n	wgblas ms	wgblas GB/s	cuBLAS ms	cuBLAS GB/s	efficiency
1024	0.0069	4.7407	0.0036	9.0220	52.5%
65536	0.0175	119.9195	0.0115	182.8061	65.6%
1048576	0.2032	165.1561	0.1985	169.0025	97.7%
16777216	3.1516	170.3478	3.3219	161.6178	105.4%

See also:

• sine.drot.js — WebGPU s-sweep benchmark script
• sine.drot.c — CUDA / cuBLAS s-sweep reference script