Performance¶

Recommendation¶

The recommended generator for single use is PCG64DXSM although SFC64 and Xoshiro256 are both excellent alternatives. Romu is a newer generator that is also very fast.

For very large scale applications – requiring 1,000+ streams – PCG64DXSM combined with a SeedSequence and spawn, SFC64 initialized using distinct Weyl increments (k), or one of the cryptography-based generators AESCounter (if you have hardware acceleration), EFIIX64, SPECK128, Philox, or HC128 if you do not) initialized with distinct keys are all excellent choices.

Unless you need backward compatibility, there are no good reasons to use any of the Mersenne Twister PRNGS: MT19937, MT64, SFMT, and DSFMT.

Timings¶

The timings below are the time in ns to produce 1 random value from a specific distribution. SFC64 is the fastest, followed closely by Xoshiro256, PCG64DXSM, JSF, and EFIIX64. The original NumPy MT19937 generator is slower since it requires 2 32-bit values to equal the output of the faster generators.

Bit Gen	Uint32	Uint64	Uniform	Expon	Normal	Gamma
Romu(variant=”trio”)	2.3	2.1	2.0	5.2	9.3	18.2
SFC64	2.3	2.5	2.4	5.1	9.6	18.7
Romu	2.5	2.5	2.4	5.0	9.5	18.8
Xoshiro256	2.5	2.5	2.5	5.2	9.8	19.2
PCG64DXSM	2.3	2.9	3.0	5.3	11.0	20.7
JSF	2.4	3.0	3.0	5.8	10.2	20.0
EFIIX64	2.5	3.0	3.0	5.4	10.5	20.4
PCG64	2.3	3.1	3.1	5.9	11.3	21.4
Xoshiro512	2.7	3.5	3.3	5.8	10.3	20.3
SFMT	2.9	3.3	3.1	6.3	11.0	20.9
LXM	2.6	3.5	3.5	6.3	11.3	21.9
PCG64(variant=”dxsm-128”)	2.8	3.4	3.5	6.1	12.5	23.1
DSFMT	3.0	4.2	2.7	7.0	12.2	21.7
MT64	2.8	4.0	4.2	6.9	12.8	23.7
JSF32	3.0	4.3	4.3	6.9	11.2	22.9
Philox(n=2, w=64)	3.2	4.7	5.1	8.1	14.7	27.4
Philox	4.0	5.9	6.1	8.8	13.6	27.0
AESCounter	4.4	6.0	5.7	9.1	14.4	27.4
MT19937	3.8	6.3	7.2	9.2	14.8	28.8
ThreeFry(n=2, w=64)	4.1	6.5	6.9	9.5	15.7	30.6
NumPy	3.0	4.6	5.8	14.4	20.1	39.8
HC128	4.1	7.2	7.2	10.5	16.6	31.5
Philox(n=4, w=32)	4.2	7.6	8.7	10.8	16.5	32.6
SPECK128	5.4	8.1	9.7	11.4	17.0	33.4
ThreeFry	5.9	9.1	9.3	12.0	16.7	34.9
ChaCha(rounds=8)	6.7	10.4	10.3	13.2	18.1	36.4
ChaCha	9.7	16.6	16.4	19.6	24.4	49.0
ThreeFry(n=4, w=32)	9.0	16.8	17.5	20.7	24.3	54.1
RDRAND	129.5	129.9	129.6	136.9	139.6	287.4

The next table presents the performance relative to NumPy’s RandomState in percentage. The overall performance is computed using a geometric mean.

Bit Gen	Uint32	Uint64	Uniform	Expon	Normal	Gamma	Overall
Romu(variant=”trio”)	128	219	282	278	215	219	217
SFC64	129	189	244	283	210	213	205
Romu	118	185	242	285	212	212	202
Xoshiro256	119	185	229	279	205	208	198
PCG64DXSM	128	163	190	271	182	192	183
JSF	126	155	193	250	197	199	182
EFIIX64	120	154	193	264	190	195	181
PCG64	128	149	184	246	178	186	175
Xoshiro512	111	131	176	249	195	196	170
SFMT	104	143	185	228	183	190	167
LXM	114	135	164	227	177	182	162
PCG64(variant=”dxsm-128”)	105	138	165	234	160	172	158
DSFMT	100	110	213	206	165	184	156
MT64	104	117	136	210	156	168	145
JSF32	99	108	134	208	178	174	145
Philox(n=2, w=64)	91	98	114	177	136	145	124
Philox	75	79	95	163	147	147	112
AESCounter	67	77	102	158	139	145	109
MT19937	78	74	80	156	135	138	105
ThreeFry(n=2, w=64)	72	71	84	152	128	130	101
HC128	72	64	80	136	121	126	96
Philox(n=4, w=32)	71	61	66	134	122	122	91
SPECK128	55	57	59	127	118	119	83
ThreeFry	51	51	62	120	120	114	80
ChaCha(rounds=8)	45	45	56	109	111	109	73
ChaCha	30	28	35	73	82	81	49
ThreeFry(n=4, w=32)	33	28	33	69	82	73	48
RDRAND	2	4	4	10	14	14	7

Note

All timings were taken using Linux on an Intel Cascade Lake (Family 6, Model 85, Stepping 7) running at 3.1GHz.