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”)	1.9	2.6	2.4	3.9	7.3	18.4
Romu	2.1	2.6	2.4	4.2	7.4	19.4
SFC64	2.1	2.6	2.5	4.2	7.6	20.6
Xoshiro256	2.3	2.8	2.7	4.4	8.0	22.1
JSF	2.2	2.9	2.7	4.6	8.1	21.5
SFMT	2.2	2.8	2.5	4.7	8.7	21.5
PCG64DXSM	2.3	3.0	2.8	4.7	9.2	23.8
PCG64	2.3	3.2	2.9	4.6	9.0	23.2
Xoshiro512	2.6	3.0	2.9	5.3	8.9	21.9
Squares	2.4	2.9	2.8	5.2	11.3	27.1
AESCounter	2.8	2.9	3.0	5.3	10.4	24.4
PCG64(variant=”dxsm-128”)	2.5	3.4	3.2	4.8	10.5	25.5
LXM	2.6	3.4	3.3	5.8	9.5	24.6
MT64	2.6	3.3	3.2	5.5	10.4	26.5
DSFMT	2.8	4.2	2.5	6.3	10.1	25.1
JSF32	2.8	3.6	3.8	5.8	9.4	23.7
MT19937	2.5	4.2	4.0	6.1	10.5	20.8
EFIIX64	2.4	3.5	3.4	9.8	12.3	24.7
Philox	3.1	4.0	4.5	6.1	10.6	28.1
Philox(n=2, w=64)	3.0	4.0	4.1	6.6	12.3	30.7
ThreeFry(n=2, w=64)	3.5	5.4	5.2	7.7	13.0	31.0
TycheOpenRand	4.3	5.5	6.2	8.3	12.6	27.5
ThreeFry	4.7	6.5	6.4	8.4	12.2	23.7
HC128	3.7	5.9	5.9	8.2	13.2	34.8
SPECK128	4.4	6.4	6.8	8.5	12.7	24.7
Philox(n=4, w=32)	3.8	6.5	6.5	9.2	13.5	32.8
ChaCha(rounds=8)	4.2	7.3	7.2	9.3	13.9	26.4
Tyche	5.4	8.2	7.8	9.8	16.3	30.7
NumPy	5.1	5.2	5.2	18.9	28.3	37.9
BlaBla	7.6	12.4	12.7	14.8	17.9	36.6
ChaCha	6.9	12.9	12.7	15.1	19.4	37.6
ThreeFry(n=4, w=32)	6.2	11.4	12.8	18.8	22.4	59.8
RDRAND	882.5	886.7	883.7	912.2	903.3	1929.9

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
Bit Gen	Uint32	Uint64	Uniform	Expon	Normal	Gamma	Overall
Romu(variant=”trio”)	267	198	217	481	389	206	276
Romu	248	203	217	455	380	196	268
SFC64	242	197	206	455	370	184	259
Xoshiro256	219	185	195	432	353	171	243
JSF	232	183	192	407	348	177	243
SFMT	228	187	207	400	323	177	242
PCG64DXSM	219	172	186	406	309	160	228
PCG64	224	162	177	414	313	163	227
Xoshiro512	195	173	177	358	319	173	221
Squares	210	179	185	364	250	140	211
AESCounter	181	179	173	354	273	155	209
PCG64(variant=”dxsm-128”)	201	154	165	391	270	149	208
LXM	193	155	159	329	298	154	204
MT64	200	157	162	343	272	143	202
DSFMT	183	125	209	298	279	151	198
JSF32	182	144	138	324	301	160	196
MT19937	201	123	128	309	269	182	191
EFIIX64	212	148	155	192	229	154	179
Philox	163	131	116	311	266	135	174
Philox(n=2, w=64)	170	131	126	285	231	123	168
ThreeFry(n=2, w=64)	145	96	101	246	217	122	145
TycheOpenRand	118	94	84	227	224	138	137
ThreeFry	109	80	81	225	232	160	135
HC128	137	88	89	229	214	109	134
SPECK128	115	82	76	222	222	154	132
Philox(n=4, w=32)	134	80	80	206	209	116	127
ChaCha(rounds=8)	122	72	73	202	203	144	125
Tyche	95	63	66	193	173	124	109
BlaBla	67	42	41	128	158	104	79
ChaCha	74	40	41	125	146	101	78
ThreeFry(n=4, w=32)	83	46	41	100	126	63	71
RDRAND	1	1	1	2	3	2	1

Note

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