Less Slow

It’s 2026, and the world still trains AI models much the way it did when I bought my first Titan V almost a decade ago. Volta had just introduced Tensor Cores, and 16-bit matrix multiplications into 32-bit accumulators were all the rage. PyTorch was only starting to catch up with TensorFlow, and Mellanox was still an independent company selling the InfiniBand cards that joined supercomputers. Since then, Nvidia has bought Mellanox and folded the network into the GPU, but most training code still treats it as an afterthought. What I believe we still don’t do enough is design software around scale-up networks. ...

I’m killing my SimSIMD project and re-launching under a new name — NumKong — StringZilla’s big brother. Around 2'000 SIMD kernels for mixed precision numerics, spread across 200'000 lines of code & docstrings, for 7 programming languages. One of the larger collections online — comparable to OpenBLAS, the default NumPy BLAS (Basic Linear Algebra Subprograms) backend (detailed comparison below). Highlights: RISC-V Vector Extensions, Intel AMX & Arm SME Tiles From Vectors to Matrices and Higher-rank Tensors From BFloat16 and Float16 to Float6 — E3M2 & E2M3 on any CPU Native Int4 & UInt4 Dot Products via Nibble Algebra Neumaier & Dot2 for higher-than-BLAS precision Ozaki Scheme for Float64 GEMMs via Float32 Tile Hardware Haversine & Vincenty for Geospatial — 5'300x faster than GeoPy Kabsch & Umeyama Mesh Alignment — 200x faster than BioPython Fused MaxSim for ColBERT — GPU-Free Late Interaction Scoring WebAssembly SIMD backend for AI Sandboxes, Edge, & Browsers C 99, C++ 23, Rust, Swift, JavaScript, GoLang, & Python 🐍 All of that tested against in-house 118-bit floating point numbers and heavily profiled for both numerical stability and speed. Here’s a preview of performance numbers for the most familiar part — GEMM (General Matrix Multiply)-like batched dot products: ...

This article is about the ugliest, but potentially most useful piece of open-source software I’ve written this year. It’s messy, because UTF-8 is messy. The world’s most widely used text encoding standard was introduced in 1989. It now covers more than 1 million characters across the majority of used writing systems, so it’s not exactly trivial to work with. The example above contains multiple confusable characters: German Eszett variants 'ß'U+00DF0x C3 9F and 'ẞ'U+1E9E0x E1 BA 9E , the Kelvin sign 'K'U+212A0x E2 84 AA and ASCII 'k'U+006B0x 6B , and Greek mu 'μ'U+03BC0x CE BC vs the micro sign 'µ'U+00B50x C2 B5 . Try guessing which is which and how they are encoded in UTF-8! ...

Ten years ago Cloudflare published the “Do the ChaCha: better mobile performance with cryptography” blog post showing “ChaCha20-Poly1305” edging out “AES-256-GCM” on phones that lacked AES acceleration. Today almost every CPU ships with wide SIMD registers and AES instructions. Apple’s A14 ₂₀₂₀, M1 ₂₀₂₀, and every successor include AES acceleration, and the same is true for most mid-range and flagship Android SoCs. So does that 2015 advice still hold in 2025? I wanted a definitive answer for the ongoing UCall rewrite, so I compared them across different AWS server SKUs with the ring Rust crate, which keeps the benchmarks reproducible while exercising the same kernels shipped in mobile TLS stacks. ...

Last summer, me, Chris Lattner, and a bunch of other people across the industry gathered together for a GPU-programming hackathon at the AGI House in San Francisco. After one too many LLM optimizations, I decided to accelerate something nobody asked for! Most elections use simple plurality voting — whoever gets the most votes wins. But there are “fairer” methods that consider ranked preferences, like the Schulze method used by Wikimedia Foundation, Debian, and pirate parties worldwide. The catch? It scales $O(n³)$ with the number of candidates. ...

AWS is the world’s largest cloud provider. It’s hard to comprehend how many billions of times per second their instances compute string hashes and SHA-256 file checksums! After releasing StringZilla v4, I spun up instances of the last 3 Graviton generations, exploring optimization opportunities across NEON, SVE, and SVE2 extensions. Graviton 2 Graviton 3 Graviton 4 Context Availability year 2020 2022 2024 Process node 7 nm, TSMC 5 nm, TSMC 3nm, TSMC Architecture Neoverse N1 Neoverse V1 Neoverse V2 Max cores 64 64 96 AWS instance family *6g *7g *8g Specifics Vector extensions NEON NEON, SVE NEON, SVE, SVE2 Encryption extensions NEON+AES NEON+AES 🤦‍♂️ NEON+AES, SVE+AES SVE vector length 128b 256b 128b 🤦‍♂️ Yes, SVE on Gravitons is tricky. More on that later. ...

To my great surprise, one of the biggest current users of my StringZilla library in the Python ecosystem is one of the world’s most widely used Image Augmentation libraries - Albumentations with over 100 million downloads on PyPI, growing by 5 million every month. Last year, Albumentations swapped parts of OpenCV - the world’s most widely used image processing library with 32 million monthly downloads in Python - for my strings library 🤯 ...

I’ve just shipped StringZilla v4, the first CUDA-capable release of my SIMD-first string processing library. Which in English means that it is now fast not only on CPUs, but also on GPUs! I’ve wanted to add ROCm-acceleration for AMD GPUs 🤦‍♂️ I’ve wanted to include a parallel multi-pattern search algorithm 🤦‍♂️ I’ve wanted to publish it back in December 2024 🤦‍♂️ So not everything went to plan, but “StringZilla 4 CUDA” is finally here, bringing 500+ GigaCUPS of edit-distance calculations in a pip install-able package, and a few more tricks up its sleeve, aimed at large-scale Information Retrieval, Databases and Datalake systems, as well as Bioinformatics workloads. All under a permissive Apache 2.0 open-source license, free for commercial use. So in this post, we’ll cover some of the most interesting parts of this release, including: ...

TL;DR: Most C++ and Rust thread-pool libraries leave significant performance on the table - often running 10× slower than OpenMP on classic fork-join workloads and micro-benchmarks. So I’ve drafted a minimal ~300-line library called Fork Union that lands within 20% of OpenMP. It does not use advanced NUMA tricks; it uses only the C++ and Rust standard libraries and has no other dependencies. Update (Sep 2025): Since the v2 release, Fork Union supports NUMA and Huge Pages, as well as tpause, wfet, and other “pro” features. Check the README for details. ...

For High-Performance Computing engineers, here’s the gist: On Intel CPUs, the vaddps instruction (vectorized float addition) executes on ports 0 and 5. The vfmadd132ps instruction (vectorized fused float multiply-add, or FMA) also executes on ports 0 and 5. On AMD CPUs, however, the vaddps instruction takes ports 2 and 3, and the vfmadd132ps instruction takes ports 0 and 1. Since FMA is equivalent to simple addition when one of the arguments is 1, we can drastically increase the throughput of addition-heavy numerical kernels. ...

I’d argue that almost every open-source developer gets an extra spark of joy when someone reads the documentation and uses their tool in a way that goes beyond the classic 101 examples. It’s a rare treat even for popular projects like JSON parsers, but if you are building high-throughput software, such as analyzing millions of network packets per second, you’ll have to dig deeper. The first thing I generally check in such libraries is the memory usage pattern and whether I can override the default memory allocator. It is the most common singleton of them all! ...

It’s 2025. Sixteen years ago, someone asked on StackOverflow how to split a string in C++. With 3000 upvotes, you might think this question has been definitively answered. However, the provided solutions can be greatly improved in terms of both flexibility and performance, yielding up to a 10x speedup. In this post, we’ll explore three better ways to split strings in C++, including a solution I briefly mentioned in 2024 as part of a longer review of the Painful Pitfalls of C++ STL Strings. ...

This blogpost is a mirror of the original post on Modular.com. Modern CPUs have an incredible superpower: super-scalar operations, made available through single instruction, multiple data (SIMD) parallel processing. Instead of doing one operation at a time, a single core can do up to 4, 8, 16, or even 32 operations in parallel. In a way, a modern CPU is like a mini GPU, able to perform a lot of simultaneous calculations. Yet, because it’s so tricky to write parallel operations, almost all that potential remains untapped, resulting in code that only does one operation at a time. ...

Set intersections are one of the standard operations in databases and search engines. They are used in: TF-IDF ranking in search engines, Table Joins in OLAP databases, Graph Algorithms. Chances are, you rely on them every day, but you may not realize that they are some of the most complex operations to accelerate with SIMD instructions. SIMD instructions make up the majority of modern Assembly instruction sets on x86 and Arm. They can yield 10x speedups, but due to their complexity, they are almost never used in production codebases. ...

Python has a straightforward syntax for positional and keyword arguments. Positional arguments are arguments passed to a function in a specific order, while keyword arguments are passed to a function by name. Surprising to most Python developers, the choice of positional vs keyword arguments can have huge implications on readability and performance. Let’s take the cdist interface as an example. It’s a function implemented in SimSIMD, mimicking SciPy, that computes all pairwise distances between two sets of points, each represented by a matrix. It accepts up to 6 arguments: ...

Downloaded over 5 Billion times, NumPy is the most popular library for numerical computing in Python. It wraps low-level HPC libraries like BLAS and LAPACK, providing a high-level interface for matrix operations. BLAS is mainly implemented in C, Fortran, or Assembly and is available for most modern chips, not just CPUs. BLAS is fast, but bindings aren’t generally free. So, how much of the BLAS performance is NumPy leaving on the table? ...

Criticizing software is easy, yet the C++ and C standard libraries have withstood the test of time admirably. Nevertheless, they are not perfect. Especially the <string>, <string_view>, and <string.h> headers. The first two alone bring in over 20,000 lines of code, slowing the compilation of every translation unit by over 100 milliseconds. Most of that code seems dated, much slower than LibC, and equally error-prone, with interfaces that are very hard to distinguish. ...

TLDR: I’ve finally finished a project that involved gathering 7 billion small molecules, each represented in SMILES notation and having fewer than 50 “heavy” non-hydrogen atoms. Those molecules were “fingerprinted”, producing 28 billion structural embeddings, using MACCS, PubChem, ECFP4, and FCFP4 techniques. These embeddings were indexed using Unum’s open-source tool USearch, to accelerate molecule search. This extensive dataset is now made available globally for free, thanks to a partnership with AWS Open Data. You can find the complete data sheet and scripts for data visualization on GitHub. ...

Experienced devs may want to skip the intro or jump immediately to the conclusions. The backbone of many foundational software systems — from compilers and interpreters to math libraries, operating systems, and database management systems — is often implemented in C and C++. These systems frequently offer Software Development Kits (SDKs) for high-level languages like Python, JavaScript, Go, C#, Java, and Rust, enabling broader accessibility. But there is a catch. Most of those SDKs are just wrappers calling your standalone application through the networking stack. That is, however, extremely slow. A good networking stack can handle over 100,000 calls per second, while most are below 10,000. ...

In this fourth article of the “Less Slow” series, I’m accelerating Unum’s open-source Vector Search primitives used by some great database and cloud providers to replace Meta’s FAISS and scale-up search in their products. This time, our focus is on the most frequent operation for these tasks - computing the the Cosine Similarity/Distance between two vectors. It’s so common, even doubling it’s performance can have a noticeable impact on applications economics. But compared to a pure Python baseline our single-threaded performance grew by a factor of 2'500x. Let’s go through optimizations one by one: ...

When our Python code is too slow, like most others we switch to C and often get 100x speed boosts, just like when we replaced SciPy distance computations with SimSIMD. But imagine going 100x faster than C code! It sounds crazy, especially for number-crunching tasks that are “data-parallel” and easy for compilers to optimize. In such spots the compiler will typically “unroll” the loop, vectorize the code, and use SIMD instructions to process multiple data elements in parallel. But I’ve found a simple case, where the GNU C Compiler (GCC 12) failed to do so, and lost to hand-written SIMD code by a factor of 119. Enough to make the friendly bull compiler cry. ...

You’ve probably heard about AI a lot this year. Lately, there’s been talk about something called Retrieval Augmented Generation (RAG). Unlike a regular chat with ChatGPT, RAG lets ChatGPT search through a database for helpful information. This makes the conversation better and the answers more on point. Usually, a Vector Search engine is used as the database. It’s good at finding similar data points in a big pile of data. These data points are often at least 256-dimensional, meaning they have many Number-s. If you use JavaScript, you might wonder whether to use the built-in Array type or the more specialized TypedArray for this job. ...

Python’s not the fastest language out there. Developers often use tools like Boost.Python and SWIG to wrap faster native C/C++ code for Python. PyBind11 is the most popular tool for the job not the quickest. NanoBind offers improvements, but when speed really matters, we turn to pure CPython C API bindings. With StringZilla, I started with PyBind11 but switched to CPython to reduce latency. The switch did demand more coding effort, moving from modern C++17 to more basic C99, but the result is a 5x lower call latency! So if you want to make your bindings lighter and better understand how CPython works under the hood - continue reading 🤗 ...

Over the years, Intel’s 512-bit Advanced Vector eXtensions (AVX-512) stirred extensive discussions. While introduced in 2014, it wasn’t until recently that CPUs began providing comprehensive support. Similarly, Arm Scalable Vector Extensions (SVE), primarily designed for Arm servers, have also started making waves only lately. The computing landscape now looks quite different with powerhouses like Intel’s Sapphire Rapids CPUs, AWS Graviton 3, and Ampere Altra entering the fray. Their arrival brings compelling advantages over the traditional AVX2 and NEON extensions: ...

A few years back, I found a simple trick in tandem with SIMD intrinsics to truly unleash the power of contemporary CPUs. I put the strstr of LibC and the std::search of the C++ Standard Templates Library to the test and hit a throughput of around 1.5 GB/s for substring search on a solitary core. Not too shabby, right? But imagine, that the memory bandwidth could theoretically reach a striking 10-15 GB/s per core. ...

Some of the most common questions in programming interviews are about strings - reversing them, splitting, joining, counting, etc. These days, having to interview more and more developers across the whole spectrum, we see how vastly the solutions, even to the most straightforward problems, differ depending on experience. Let’s imagine a test with the following constraints: You must find the first occurrence of every unique string in a non-empty array. You are only allowed to use the standard library, no other dependencies. You have 20 minutes. In Python, the solution may look like this: ...

There are only two kinds of languages: the ones people complain about and the ones nobody uses. – Bjarne Stroustrup, creator of C++. Very few consider C++ attractive, and only some people think it’s easy. Choosing it for a project generally means you care about the performance of your code. And rightly so! Today, machines can process hundreds of Gigabytes per second, and we, as developers, should all learn to saturate those capabilities. So, let’s look into a few simple code snippets and familiarize ourselves with Google Benchmark (GB) - the most famous library in the space using incrementally complex examples. ...

A bit of history. Not so long ago, we tried to use GPU acceleration from Python. We benchmarked NumPy vs CuPy in the most common number-crunching tasks. We took the highest-end desktop CPU and the highest-end desktop GPU and put them to the test. The GPU, expectedly, won, but not just in Matrix Multiplications. Sorting arrays, finding medians, and even simple accumulation was vastly faster. So we implemented multiple algorithms for parallel reductions in C++ and CUDA, just to compare efficiency. CUDA was obviously harder, than using std::accumulate, but there is a shortcut: thrust::reduce. ...

GPU acceleration can be trivial for Python users. Follow CUDA installation steps carefully, replace import numpy as np with import cupy as np, and you will often get the 100x performance boosts without breaking a sweat. Every time you write magical one-liners, remember a systems engineer is making your dreams come true. A couple of years ago, when I was giving a talk on the breadth of GPGPU technologies, I published a repo. A repo with various cool GPGPU staff in Vulkan, Halide, and most importantly, 8x implementations of std::accumulate in OpenCL. I know what you are thinking: ...

This will be a story about many things: about computers, about their (memory) speed limits, about very specific workloads that can push computers to those limits and the subtle differences in Hash-Tables (HT) designs. But before we get in, here is a glimpse of what we are about to see. A friendly warning, the following article contains many technical terms and is intended for somewhat technical and hopefully curious readers. ...

David Patterson had recently mentioned that (rephrasing): The programmers may benefit from using complex instruction sets directly, but it is increasingly challenging for compilers to automatically generate them in the right spots. In the last 3-4 years I gave a bunch of talks on the intricacies of SIMD programming, highlighting the divergence in hardware and software design in the past ten years. Chips are becoming bigger and more complicated to add more functionality, but the general-purpose compilers like GCC, LLVM, MSVC and ICC cannot keep up with the pace. Hardly any developer codes in Assembly today, hoping that the compiler will do the heavy lifting. ...