Question: Architecture-specific performance differences between blake3 and blake2

[tasleson]

Summary

I’ve been evaluating the Rust blake3 crate for potential use. Data, code, and benchmarks are available here.

Results

• x86: blake3 is consistently faster than blake2.
• ppcle / s390x / aarch64: performance is generally slower than blake2.
  • Rayon parallelism sometimes improves results.
  • In some cases, performance is still worse (example).

Questions

• Is this expected behavior on non-x86 architectures (e.g., SIMD gaps, missing intrinsics)?
• Or is my sample code / benchmarking harness flawed?
• Are there recommended tuning options or build flags for ppcle, s390x, and aarch64?

[tasleson]

How I run the test ...

$ cargo build --release --features "b3rayon"
$ target/release/blake-bench --markdown --include-parallel > result.md

[sneves]

The blake2s_256 timings seem to be actually using BLAKE2b underneath: https://github.com/tasleson/misc-tools/blob/378b8e9c6052ca8797d18f19649f686325c343f6/blake-bench/src/main.rs#L192-L193

[tasleson]

The blake2s_256 timings seem to be actually using BLAKE2b underneath: https://github.com/tasleson/misc-tools/blob/378b8e9c6052ca8797d18f19649f686325c343f6/blake-bench/src/main.rs#L192-L193

You're correct, thanks for the catch!