Multiqubit ops GPU performance

[stavros11]

Following our discussion, this is the GPU equivalent of #51 where we can discuss various interesting findings regarding the multiqubit kernel performance on GPU and particularly how it compares to qiskit. Here are some benchmark results:

Simulation time qibo/qiskit - RTX A6000 - double precision

Simulation time qibo/qiskit - RTX A6000 - single precision

In single precision we are much faster while in double precision we are mostly equivalent apart from a specific area. Here are the exact times for this interesting area:

simulation times - nqubits=15

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.00068	0.00171	0.00057	0.00115
2	0.00105	0.00177	0.00066	0.00184
3	0.00270	0.00184	0.00112	0.00175
4	0.00695	0.00169	0.00125	0.00242
5	0.02246	0.00259	0.00246	0.00350
6	0.03912	0.01039	0.00538	0.00637
7	0.06972	0.02882	0.00678	0.02115
8	0.12242	0.11884	0.01668	0.10134
9	0.25151	0.60537	0.05127	0.57179
10	0.77184	3.72334	0.30835	3.54705

simulation times - nqubits=16

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.00071	0.00142	0.00061	0.00132
2	0.00108	0.00193	0.00069	0.00191
3	0.00286	0.00210	0.00120	0.00183
4	0.00756	0.00758	0.00135	0.00232
5	0.02427	0.00859	0.00269	0.00351
6	0.08424	0.01121	0.00997	0.00782
7	0.15274	0.03605	0.01274	0.02318
8	0.27393	0.11994	0.02289	0.11415
9	0.52561	0.72564	0.06245	0.65283
10	1.15000	4.27385	0.36389	4.22007

simulation times - nqubits=17

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.00076	0.00213	0.00065	0.00203
2	0.00121	0.00233	0.00074	0.00170
3	0.00307	0.00792	0.00129	0.00286
4	0.00813	0.00277	0.00146	0.00361
5	0.02625	0.00923	0.00291	0.00575
6	0.09226	0.01316	0.01078	0.01105
7	0.33222	0.04734	0.02657	0.03166
8	0.60661	0.17166	0.04665	0.13848
9	1.09221	0.74349	0.08526	0.75295
10	2.17680	4.90263	0.41825	4.97029

simulation times - nqubits=18

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.00092	0.00196	0.00071	0.00226
2	0.00122	0.00227	0.00080	0.00225
3	0.00329	0.00906	0.00134	0.00408
4	0.00871	0.00968	0.00154	0.00623
5	0.02826	0.01136	0.00312	0.01024
6	0.09959	0.02241	0.01180	0.01939
7	0.36258	0.06104	0.02912	0.04782
8	1.32488	0.23768	0.10081	0.17849
9	2.42261	0.87334	0.19749	0.88229
10	4.36327	5.55920	0.57205	5.60554

simulation times - nqubits=19

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.00127	0.01038	0.00077	0.00300
2	0.00191	0.00469	0.00089	0.00329
3	0.00357	0.01160	0.00145	0.00656
4	0.00945	0.02012	0.00165	0.01049
5	0.03066	0.01703	0.00339	0.01806
6	0.10760	0.03546	0.01305	0.03311
7	0.39375	0.08286	0.03169	0.07409
8	1.44949	0.23880	0.11003	0.23198
9	5.31427	0.98861	0.39713	1.04869
10	9.69975	6.31365	0.82073	6.39574

simulation times - nqubits=20

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.00182	0.01452	0.00123	0.00561
2	0.00312	0.01466	0.00133	0.00602
3	0.00642	0.01459	0.00187	0.01225
4	0.01038	0.03097	0.00184	0.02110
5	0.03337	0.03782	0.00392	0.03531
6	0.11569	0.05954	0.01418	0.06470
7	0.42412	0.12040	0.03445	0.13335
8	1.57073	0.24070	0.11950	0.34332
9	5.79949	1.12296	0.43513	1.31031
10	21.42194	6.96956	1.59364	7.14927

simulation times - nqubits=21

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.00297	0.04279	0.00183	0.01087
2	0.00506	0.04736	0.00190	0.00970
3	0.01134	0.04779	0.00257	0.02269
4	0.02009	0.04995	0.00272	0.04079
5	0.03781	0.07061	0.00496	0.06971
6	0.12480	0.11806	0.03517	0.12507
7	0.45471	0.19977	0.05067	0.24671
8	1.69274	0.37102	0.13072	0.56319
9	6.32958	1.37186	0.47240	1.76927
10	23.37317	7.95625	1.75441	8.69835

simulation times - nqubits=22

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.00529	0.04282	0.00304	0.02550
2	0.00885	0.04292	0.00307	0.02752
3	0.01958	0.04581	0.00373	0.05389
4	0.03942	0.05852	0.00431	0.08455
5	0.07661	0.08379	0.00905	0.14375
6	0.13771	0.13358	0.07412	0.25818
7	0.48743	0.24396	0.10509	0.48839
8	1.82252	0.55038	0.19315	1.01881
9	6.81780	1.78645	0.52388	2.68159
10	25.32039	9.04470	1.91643	10.59537

simulation times - nqubits=23

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.01015	0.08001	0.00550	0.04743
2	0.01712	0.08031	0.00543	0.04986
3	0.03556	0.08724	0.00609	0.10628
4	0.07037	0.11227	0.00700	0.17147
5	0.15528	0.16596	0.01858	0.29656
6	0.29346	0.26994	0.15351	0.53666
7	0.52695	0.48206	0.22443	1.00201
8	1.96243	0.99488	0.40818	1.99313
9	7.36424	2.65864	0.75142	4.56154
10	27.27548	11.00970	2.12359	14.71918

simulation times - nqubits=24

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.02022	0.18080	0.01053	0.08798
2	0.03352	0.18370	0.01043	0.09721
3	0.07010	0.19461	0.01097	0.21545
4	0.13568	0.25227	0.01237	0.35247
5	0.28303	0.36481	0.03299	0.61607
6	0.61949	0.58082	0.32582	1.12560
7	1.16325	1.00706	0.47051	2.11143
8	2.11352	1.91901	0.86623	4.03778
9	7.86494	4.39171	1.59513	8.44242
10	29.23039	14.76853	3.00823	22.39044

simulation times - nqubits=25

ntargets	qibo double (sec)	qiskit double (sec)	qibo single (sec)	qiskit single (sec)
1	0.04113	0.36355	0.02106	0.18211
2	0.06760	0.37848	0.02062	0.20346
3	0.14073	0.39008	0.02098	0.44870
4	0.27040	0.51104	0.02358	0.73671
5	0.54744	0.75356	0.06350	1.30015
6	1.14574	1.21546	0.67175	2.36441
7	2.47845	2.06826	1.00970	4.39131
8	4.59437	4.11890	1.82457	8.32751
9	8.45306	9.00420	3.41029	16.53761
10	31.44260	24.93261	6.36849	38.09419

It appears that qiskit has an issue here as in some cases their single precision is significantly slower than their double. On the other hand, qibo's single precision is extremely faster than double.

Below are some numbers from the DGX, where the situation is completely different and more in line with what we observed on CPU in #51:

Simulation time qibo/qiskit - V100 - double precision

simulation times - nqubits=24 - V100

ntargets	qibo double (sec)	qiskit double (sec)
3	0.42339	0.33970
4	0.62654	0.33986
5	1.00678	0.33991
6	2.19785	0.34998
7	3.63109	0.45986
8	6.53215	0.76321
9	11.95559	2.57002
10	22.17884	15.57758

simulation times - nqubits=25 - V100

ntargets	qibo double (sec)	qiskit double (sec)
3	0.80510	0.67170
4	1.23495	0.67187
5	2.09212	0.67181
6	4.59987	0.67553
7	7.65031	0.88789
8	13.84615	1.25915
9	25.54963	3.41262
10	47.12159	19.35345

simulation times - nqubits=26 - V100

ntargets	qibo double (sec)	qiskit double (sec)
3	1.65588	1.33589
4	2.56164	1.33560
5	4.35132	1.33077
6	9.55329	1.35739
7	16.05355	1.74887
8	29.17296	2.29261
9	53.89957	5.04608
10	100.10543	26.62849

simulation times - nqubits=27 - V100

ntargets	qibo double (sec)	qiskit double (sec)
3	3.41055	2.65044
4	5.32581	2.63898
5	9.09922	2.63975
6	20.06714	2.64693
7	33.64388	3.46601
8	61.26683	4.29842
9	113.86667	8.26060
10	212.27658	40.63561

simulation times - nqubits=28 - V100

ntargets	qibo double (sec)	qiskit double (sec)
3	7.15363	5.24736
4	11.12589	5.22061
5	18.98035	5.25438
6	41.96390	5.26979
7	70.22372	6.85323
8	128.53272	8.41681
9	239.46607	14.80653
10	448.56989	70.21813

NOTE: I just realized that the benchmark script we have been using does not use state.numpy() so the final state is not transferred from GPU to CPU for qibo, something that probably happens for qiskit which returns a numpy array. If we include the transfer, qibo's results may be worse (not sure how much), but I believe some of above observations will still hold. For example the strange fact that qiskit single is slower than qiskit double remains, but is not really our problem to solve.

[scarrazza]

Closing, obsolete results.