Fearless SIMD 1.0 Is Stable, but Rust's std::simd Is Still Nightly

Fearless SIMD reached version 1.0 on Sept. 21, giving Rust programmers a stable, safe way to write SIMD code, the instructions that apply one operation to several values at once. Rust's own std::simd is still nightly-only, according to a 31-minute survey published Sept. 25 by Sergey "Shnatsel" Davidoff.

The release notes set the terms: a minimum Rust of 1.89, and this promise: "Starting with v1.0, the latest Fearless SIMD version for each MSRV will receive security backports for at least three years after that Rust version was released." The crate is maintained by the Linebender organisation and dual-licensed Apache 2.0 or MIT, and its repository showed 617 stars when The Terminal read it.

What 1.0 supports

The release covers native implementations for SSE2, SSE4.2, AVX2 and AVX-512 on x86, NEON on ARM, and SIMD128 plus relaxed-simd on WebAssembly, with a scalar fallback. Version 1.0 also adds an optional #[simd] attribute for function-level SIMD and reductions such as reduce_sum and reduce_max.

The five libraries Davidoff compares

Davidoff's survey sorts the options into automatic vectorization, portable abstractions and raw intrinsics. His comparison of the abstractions:

Library	Status in the survey	Main weakness he names
std::simd	Nightly	API instability; poor float performance
fearless_simd	1.0 stable	Trigonometry gap; boilerplate
wide	1.0 stable	Does not combine with multiversioning on x86
pulp	Production, powers faer	Verbose multiversioning; thin docs
macerator	Production, powers burn	Minimal documentation

Multiversioning means shipping several code paths and picking one at run time by CPU. It matters because of the hardware numbers below.

Why AVX2 is still the floor

AVX2 shipped on CPUs in 2013, yet Davidoff cites the Firefox hardware survey for the figure that 15 percent of x86 systems lack it in 2026. He says early AVX-512 chips downclocked under load and that only Ice Lake, from 2019 on, made it worthwhile.

On ARM his verdict is kinder: 64-bit NEON gets right what AVX2 got wrong, being mandatory, 128 bits wide and complete enough to need no multiversioning. His view of RISC-V for 2026 is that it is essentially irrelevant to this problem.

What changed in the compiler

Davidoff credits Rust 1.98 with stabilising algebraic float operations such as algebraic_add(), which let the compiler reorder floating-point maths and so vectorize loops it would otherwise leave alone. Since Rust 1.87, intrinsics that touch no memory can be called without an unsafe block. His compiler wishlist starts with min_generic_const_args and the Struct Target Features RFC.

The gap he flags most often is trigonometry. The 1.0 release does not close that, so the first thing to check in a maths-heavy crate is whether a stable trig routine exists before choosing a library.

Where NEON matters in practice

Davidoff's point about NEON being mandatory on 64-bit ARM has a practical consequence for phone and laptop software: code that targets it does not need a run-time dispatch step. Our piece on the postmarketOS rebrand as Nura covers a Linux project that runs on ARM phones, where that baseline applies. For a different class of software that leans on vector maths, see our coverage of PrismML's ternary Bonsai 2 model.

Rust's 1.89 floor is the number to plan around, since it is the version the 1.0 backport policy is anchored to. Check your own minimum supported Rust version against it before adopting the crate.