Scientific Computing in Rust Monthly #18
Welcome to this month's issue of Scientific Computing in Rust Monthly.
Scientific Computing in Rust Monthly can be viewed online at scientificcomputing.rs/monthly, you can sign up to receive an email copy at mailinglist.scientificcomputing.rs, or you can subscribe using the RSS feed at scientificcomputing.rs/monthly/rss.xml.
Items can be added to the next issue of the newsletter by opening a pull request to the Scientific Computing in Rust Monthly repository on GitHub. You can find more information about adding items here.
Crate of the month: hotpath
crates.io ~ docs.rs ~ source ~ www
Hotpath-rs is an easy-to-configure Rust performance profiler.
It shows exactly where your code spends time, burns CPU, and allocates memory.
Hotpath-rs helps you distinguish between functions that are slow
because they wait on I/O and those that are CPU-intensive.
Trying hotpath-rs out in self-profiling mode is as easy as typing
ssh demo.hotpath.rs.
New releases
burn v0.21.0
After five pre-releases over the last three months, burn v0.21.0
was recently released.
The gains span distributed workflows for training large models
all the way down to small-model inference,
where the reduced framework overhead becomes especially noticeable.
Full release notes can be found
on the GitHub release page.
cuda-oxide v0.1.0
Cuda-oxide is an experimental rustc backend
for compiling CUDA GPU kernels written in pure Rust to PTX.
It supports single-source Rust GPU programs where host
and device code live together and are built through cargo oxide.
delaunay v0.7.8
Delaunay v0.7.8 provides D-dimensional Delaunay triangulations and convex hulls (2D through 5D explicitly tested) with PL-manifold (default) or pseudomanifold guarantees on finite point sets with Euclidean and toroidal global topologies. It uses exact predicates, Simulation of Simplicity, and Hilbert curves for robust degeneracy handling, deterministic insertion ordering, and efficient spatial indexing. It provides a 4-level validation hierarchy covering elements, data structure validity, manifold topology, and Delaunay adherence. It supports Pachner moves up to D=5 via bistellar flips, vertex insertion/deletion, and bounded flip/rebuild conversion of non-Delaunay triangulations. Auxiliary data may be stored in vertices and simplices, with external secondary maps for algorithm use, and the data structure is serializable/deserializable. Written in safe Rust with no unsafe code, it is inspired by CGAL and Spade and aims to be a lightweight, rigorous Rust-native option for workflows needing explicit topology settings, validation levels, deterministic construction controls, and repair behavior.
numra v0.1.4
numra is the umbrella crate of the Numra workspace,
which consists of twenty native-Rust crates such as
differential equations, optimization, sensitivity analysis,
uncertainty quantification, linear algebra, and more.
Dense linear algebra builds on faer;
the rest is native Rust with no FFI to C or FORTRAN.
pluot v0.1.0
Pluot is a crate intended to facilitate the development of visualizations that can be rendered from multiple programming languages. This currently includes Rust, Python, and JavaScript. Further, Pluot aims to support both static and interactive use cases. For some demonstrations, see also the pluot website.
rayon v1.12.0
This new version of rayon contains a bug fix in parallel Range<char>
and added a new method ParallelSlice::par_array_windows.
Details can be found
in the release notes.
Events
Scientific Computing in Rust, 8-10 July 2026, held virutally
This year's scientific computing in Rust workshop is taking place on 8-10 July. The deadline for talk submissions is Friday 5 June: you can submit your talk at scientificcomputing.rs/2026/submit-talk.
Publications
Keller Mark S. and Gehlenborg Nils, Pluot: Towards 'write once, run everywhere' visualization software, arXiv 2605.14118
This manuscript describes Pluot, a crate to facilitate write-once, render in many places visualizations.
Pastel Arthur, 5× faster fast_blur in image-rs
An interesting algorithm discussion on a new fast_blur implementation in image-rs.
The article discusses the algorithm itself, profiling of the previous implementation,
as well as the design choices in the new version.
Especially interesting is the analysis and discussion of how fast various CPU instructions
are in comparison with each other.
