Longer is faster, on one scale for every machine. The full bar hashes nonstop on every core; the darker part is one message of 1 KiB on one thread. The bracket: how many times faster the fastest ran, in its colour. Point at a bar for figures.
A column is a machine, a row a use case. In each cell the fastest bar fills the width; the bracket says by how much it leads.
| Intel Core i7-12700K20 CPUs, Linux · 2026-10-23BLAKE3 servil 082afa3 · bench-hashes 0.17.0 | AMD EPYC 925548 CPUs, Linux · 2026-10-23BLAKE3 servil 082afa3 · bench-hashes 0.17.0 | Apple M4 Max16 CPUs, macOS · 2026-10-24BLAKE3 servil b94e2fa · bench-hashes 0.17.0 | VM on aarch6416 CPUs, Linux · 2026-10-24BLAKE3 servil b94e2fa · bench-hashes 0.17.0 | |
|---|---|---|---|---|
| 64 B | 0.079 GB/s | 0.14 GB/s | 0.11 GB/s | 0.047 GB/s |
| 1 KiB | 0.68 GB/s | 0.84 GB/s | 1.0 GB/s | 0.60 GB/s |
| 1 MiB, one thread | 3.6 GB/s | 7.9 GB/s | 6.0 GB/s | 5.5 GB/s |
| 128 MiB, every core | 13 GB/s | 30 GB/s | 28 GB/s | 24 GB/s |
| 64 KiB messages, one after another | 21 GB/s | 11 GB/s | 30 GB/s | 21 GB/s |
| 1024 messages of 64 B, batched | 3.4 GB/s | 6.3 GB/s | 4.7 GB/s | 4.2 GB/s |
| 1000 files, b3sum | 14 GB/s | 19 GB/s | 6.7 GB/s | 14 GB/s |
BLAKE3 official is the blake3 crate, the reference implementation; BLAKE3 servil is this fork of it, with the same digests. SHA-256 is the fastest implementation measured on each machine (the sha2 crate, ring, or the operating system's CommonCrypto); the tooltip names which. sha1sum and md5sum, drawn in grey, compute SHA-1 and MD5: both are broken as cryptographic hashes (collisions are cheap to make), and they appear here only because many people still run them as checksums; a comparison with them is a comparison with the tool in hand, not with a safe alternative.