SHISHUA – The Fastest PRNG In The World
The announcement and explanation blog post is here.
Note: please do not use this for cryptographic purposes. If you need security, a recommended option is ChaCha20.
Implementations:
Comparison
| Name | Intel Performance | ARM Performance | Quality | Seed correlation |
|---|---|---|---|---|
| SHISHUA | 52.93 GB/s | 8.78 GB/s | >32 TiB | >32 TiB |
| xoshiro256+x8 | 40.18 GB/s | 3.69 GB/s | 1 KiB | 0 KiB |
| RomuTrio | 8.62 GB/s | 4.76 GB/s | >32 TiB | 1 KiB |
| xoshiro256+ | 7.88 GB/s | 4.31 GB/s | 512 MiB | 1 KiB |
| wyrand | 7.00 GB/s | 2.53 GB/s | >32 TiB | 32 KiB |
| Lehmer128 | 6.14 GB/s | 1.83 GB/s | >32 TiB | 1 KiB |
| ChaCha8 | 6.28 GB/s | 1.72 GB/s | >32 TiB? | >32 TiB? |
| RC4 | 0.35 GB/s | 0.15 GB/s | 1 TiB | 1 KiB |
- Performance: in number of CPU cycles spent per byte generated, on N2 GCP instances. On N2D (AMD), the order is the same.
- Quality: level at which it fails PractRand. We show a
>if it did not fail. We put a question mark if we have not proved it. - Seed correlation: PractRand on interleaving of bytes from eight streams with seeds 1, 2, 4, 8, 16, 32, 64, 128. We use PractRand with folding 2 and expanded tests.
On the subject of seed correlations, the ./bin/sample-seed-fingerprints.sh program
highlights additional anomalies:
- RC4 has weak seeds with heavy artefacts visible in the fingerprints.
- RomuTrio's fingerprints start with the same first character for all seeds.
Commands
make: build./shishua.make test: run performance tests, PractRand, and BigCrush on SHISHUA.make test/benchmark-seed: run seed correlation test.make test/benchmark-perf: run performance comparison locally.make benchmark-intel: run performance comparison on a GCP Intel chip.make benchmark-amd: run performance comparison on a GCP AMD chip.make benchmark-arm: run performance comparison on an AWS Graviton ARM chip.
The SHISHUA and SHISHUA-half are under the CC0 license.