Performance Random Performance

Random Number Generator Performance Benchmarks

State repair for a generator restored from JSON or protobuf happens in its constructor or the shared after-deserialization callback. This includes each generator's state, the shared bit/byte reservoirs and GUID scratch buffer. Repair is not repeated in a draw method, so malformed/default serialized state is repaired before the first draw without adding a guard to every later draw. PhotonSpin's one-time warmup priming also happens there, leaving only its block-boundary check in the draw path.

How the Speed column is measured

The ops/s columns are each generator measured on its own, one after another. That answers "how fast is this generator here" and it is the wrong instrument for "which generator is faster": the roster takes minutes to walk, and anything that changed on the machine in between lands on whichever generator was being measured at the time.

The Speed column is measured differently. Every generator is compared against IllusionFlow (what PRNG.Instance returns) in a fixed ABBABAAB batch, so each of the four readings that make up a ratio sits next to the reading it is divided by, and both generators occupy the same mean position in the batch. A drift that is linear across the batch cancels rather than being attributed to one side. The four raw readings per side are kept, and their spread is what decides whether the ratio is worth publishing: if the machine moved more than 3% between adjacent cycles, that generator falls back to the un-paired number and the run says so.

So a Speed bucket is a claim about the generators; an ops/s figure is a claim about this machine on that day.

Two batteries, because they ask different questions

PractRand streams until something fails and reports the depth: a generator is "clean through 8GB" or it is not. TestU01 SmallCrush runs fifteen fixed statistics and reports a p-value for each. A generator can be clean at 8GB of PractRand and still land a decisive p-value here, so neither stands in for the other.

Both run against the same byte stream, from the same host, so a difference between them is a difference between the batteries rather than between two ways of producing bytes.

Reading a SmallCrush result takes one piece of context: with fifteen statistics, a perfectly good generator lands one p-value outside [0.001, 0.9990] roughly one run in seven. The threshold that separates noise from signal is not close: the recorded weak control reports eps (below 1e-300), while noise sits around 1e-4. Anything below 1e-10 is treated as a failure; anything above it is reported and ignored. IllusionFlow produced a single 7.2e-4 at the manifest seed and was clean on two other seeds, which is what that rule is for.

Whole-inventory SmallCrush, 2026-08-26, seed 00010203-0405-0607-0809-0a0b0c0d0e0f, 908 MB and 6.3 s of CPU per generator:

• Every generator rated Good or better passed.
• Four recorded-weak generators failed decisively, most of their statistics at eps: LinearCongruentialGenerator, WDoomRandom, WaveSplatRandom, XorShiftRandom.
• Three recorded-weak generators passed: DotNetRandom, SquirrelRandom, SystemRandom. SmallCrush is the shallower instrument, so that is inconclusive rather than a contradiction of their rating.

For statistical batteries, the repository's Generator~/WallstopStudios.UnityHelpers.RandomQuality host emits a reproducible little-endian byte stream from an explicit generator, GUID seed and byte count. Long PractRand/TestU01 runs belong in scheduled reporting with pinned tools and expected weak-generator failures; they are not suitable as nondeterministic pull-request gates.

The NextUlong, NextLong and NextDouble columns below predate the one-advance-per-64-bit-draw change to BlastCircuitRandom, RomuDuo, SplitMix64, WyRandom and Xoshiro256StarStar, which is measured at 2.49x on Unity 6000.4.6f1 (Mono). They refresh the next time RandomPerformanceTests.Benchmark runs.

Summary (fastest first)

Random	NextUint (ops/s)	Speed	Quality	Notes
LinearCongruentialGenerator	1,323,700,000	Fastest	Poor	Numerical Recipes 'quick and dirty' LCG (a=1664525, c=1013904223, m=2^32) returning the raw state, so bit k has period only 2^(k+1); measured linear complexity of bit k is 2^k+1, and bit 0 simply alternates. Cosmetic use only.
WaveSplatRandom	1,283,400,000	Fastest	Experimental	Single-word chaotic generator; author notes period 2^64 but provides no formal test results, so treat it as experimental.
BlastCircuitRandom	1,054,900,000	Very Fast	Good	Chaotic ARX mixer rather than a proven statistically optimal generator. Verified here: PractRand 0.95 clean through 8GB, the depth at which SystemRandom fails.
SplitMix64	1,052,300,000	Very Fast	Very Good	Well-known SplitMix64 mixer; passes TestU01 BigCrush and PractRand up to large data sizes in literature. Vigna 2014
FlurryBurstRandom	923,200,000	Fast	Excellent	Six-word ARX-style generator tuned for all-around use. Verified here: PractRand 0.95 clean through 8GB, the depth at which SystemRandom fails. The author reports TestU01 BigCrush passes; that run cannot be checked because the upstream repository is offline.
PcgRandom	897,900,000	Fast	Excellent	PCG XSH RR 64/32 variant; passes TestU01 BigCrush and PractRand in published results. O'Neill 2014
XoroShiroRandom	754,700,000	Fast	Good	xoroshiro128+, returning the high 32 bits, the half its authors recommend. The discarded low half is linear (bit 0 has linear complexity exactly 128); no output bit of the returned half is. A 64-bit draw costs two state advances, because a + scrambler has no strong 64-bit word to return. Blackman & Vigna 2018
IllusionFlow	754,500,000	Fast	Excellent	Five-word rotate/xor/add generator driven by a 32-bit Weyl counter, and the generator PRNG.Instance returns. Verified here: PractRand 0.95 clean through 8GB, the depth at which SystemRandom fails. The author reports 64GB; that run cannot be checked because the upstream repository is offline.
RomuDuo	750,100,000	Fast	Good	Published romuDuo: the ROMU multiplier with the ROTL(y,36) + ROTL(y,15) - x update. NextUint returns the low 32 bits of the 64-bit word. Throughput figures in this table predate that update rule and have not been re-measured. Overton 2020
StormDropRandom	705,200,000	Moderate	Excellent	Large-state ARX generator over a 1024-word (4 KB) ring buffer with two 32-bit control words. Verified here: PractRand 0.95 clean through 8GB, the depth at which SystemRandom fails. The author's own results cannot be checked because the upstream repository is offline.
XorShiftRandom	602,100,000	Moderate	Fair	Classic 32-bit xorshift; known to fail portions of TestU01 and PractRand, acceptable for lightweight effects only. Marsaglia 2003
WyRandom	440,500,000	Slow	Very Good	Wyhash-based generator; published testing shows it clears BigCrush/PractRand with wide seed coverage. Wang Yi 2019
SquirrelRandom	414,000,000	Slow	Fair	Hash-based generator built on Squirrel3. Measured: fails PractRand 0.95 FPF-14+6/16 at 1GB, reproducibly across four seeds. Good equidistribution for the table lookups it was designed for; not a general-purpose stream. Squirrel Eiserloh
PhotonSpinRandom	261,100,000	Very Slow	Excellent	SHISHUA-inspired generator. Verified here: PractRand 0.95 clean through 8GB, the depth at which SystemRandom fails. The author reports 128GB; that run cannot be checked because the upstream repository is offline.
UnityRandom	87,600,000	Very Slow	Fair	Mirrors UnityEngine.Random, documented by Unity as Xorshift 128; suitable for legacy compatibility but not high-stakes simulation. UnityEngine.Random
SystemRandom	64,700,000	Very Slow	Poor	Reimplements the classic .NET System.Random algorithm (Knuth subtractive lagged-Fibonacci, mod 2^31-1) so its sequence is serializable and stable across runtimes. Fails modern statistical batteries. System.Random considered harmful
DotNetRandom	55,900,000	Very Slow	Poor	Wraps System.Random, which on Mono and for seeded .NET Core is Knuth's subtractive lagged-Fibonacci generator (mod 2^31-1), not an LCG. The sequence is runtime-dependent, so do not rely on it for cross-platform determinism. System.Random considered harmful

Detailed Metrics

Random	NextBool	Next	NextUint	NextFloat	NextDouble	NextUint (Range)	NextInt (Range)
LinearCongruentialGenerator	785,200,000	538,800,000	1,323,700,000	184,300,000	296,300,000	582,900,000	498,700,000
WaveSplatRandom	787,300,000	528,500,000	1,283,400,000	184,300,000	297,900,000	529,000,000	458,200,000
BlastCircuitRandom	788,200,000	537,400,000	1,054,900,000	183,800,000	293,200,000	479,700,000	421,200,000
SplitMix64	795,900,000	537,200,000	1,052,300,000	184,400,000	297,700,000	482,500,000	439,100,000
FlurryBurstRandom	767,800,000	526,500,000	923,200,000	182,300,000	293,100,000	449,800,000	404,100,000
PcgRandom	249,800,000	527,300,000	897,900,000	181,000,000	274,700,000	447,700,000	405,000,000
XoroShiroRandom	761,000,000	359,000,000	754,700,000	157,900,000	192,400,000	422,100,000	376,900,000
IllusionFlow	779,300,000	529,200,000	754,500,000	183,000,000	281,600,000	439,200,000	390,000,000
RomuDuo	784,600,000	359,400,000	750,100,000	156,100,000	188,700,000	437,700,000	384,600,000
StormDropRandom	776,600,000	523,600,000	705,200,000	180,600,000	252,400,000	393,500,000	356,700,000
XorShiftRandom	783,300,000	534,100,000	602,100,000	184,100,000	283,200,000	474,500,000	383,500,000
WyRandom	749,500,000	363,000,000	440,500,000	159,000,000	185,300,000	290,000,000	277,800,000
SquirrelRandom	756,300,000	382,700,000	414,000,000	158,800,000	197,800,000	355,500,000	309,500,000
PhotonSpinRandom	713,300,000	220,800,000	261,100,000	119,900,000	115,800,000	217,600,000	214,200,000
UnityRandom	628,900,000	76,700,000	87,600,000	59,700,000	38,800,000	81,800,000	81,900,000
SystemRandom	146,200,000	145,700,000	64,700,000	131,300,000	138,600,000	58,700,000	57,800,000
DotNetRandom	544,600,000	53,100,000	55,900,000	44,700,000	26,700,000	53,400,000	51,700,000

Generators added since the last benchmark run

The tables above are rewritten only by a benchmark run, so a generator added since the last one is absent until the .github/workflows/unity-benchmarks.yml workflow next runs. Absence here says nothing about quality: statistical standing is measured separately, by the bit-plane linearity gate on every pull request and by the scheduled PractRand battery. See Random Generators for the current ratings.

The battery runs both stream widths. NextUlong is no longer NextUint rearranged: five generators answer a 64-bit draw from one raw word, so half of it reaches a caller only through NextDouble, NextLong and NextUlong(max) and appears in no 32-bit draw. Even the generators that do build NextUlong from two NextUint draws pack them high-word-first and write little-endian, so their 64-bit stream is the 32-bit one with each adjacent word pair swapped. SystemRandom is the proof that this is not a redundant measurement: it fails the 32-bit battery at exactly 8GB and is clean through 8GB at 64-bit. Every "clean through 8GB" above is the 32-bit figure; the 64-bit outcomes are recorded per generator in scripts/random-quality/expected-outcomes.json.

Refreshing these numbers

Run RandomPerformanceTests.Benchmark from Unity's Test Runner, or let the weekly Unity Benchmarks workflow do it. Both rewrite the tables in place.