Ash's Blog

Around 2'000 SIMD kernels for mixed-precision BLAS-like numerics — dot products, batched GEMMs, distances, geospatial, ColBERT MaxSim, and mesh alignment — from Float6 to Float118, leveraging RISC-V, Intel AMX, Arm SME, and WebAssembly Relaxed SIMD, in 7…

ICU gets Unicode right and pays for it.

AES-256-GCM now beats ChaCha20-Poly1305 by up to 3x on every modern CPU with hardware acceleration, reversing the 2015 mobile performance advice.

Taking the computationally expensive Schulze voting method from theory to practice on GPUs — exploring parallel algorithms, hardware optimizations, and why Mojo surprised me.

SVE2 AES on AWS Graviton 4 delivers 2x faster string hashing than NEON, but SVE's variable-length promise fell flat: Graviton 4 regressed to 128-bit vectors.

Like Ptolemaic astronomers stacking circles to predict planetary motion, we stack transformer layers to approximate intelligence—accurate yet awaiting our Kepler moment.

StringZilla's SIMD-optimized Look-Up Tables beat OpenCV by 4x on image processing, proving string manipulation techniques excel beyond text.

StringZilla v4 brings CUDA acceleration for string processing: 109x faster than Nvidia's CuDF on edit distances, plus 52-bit MinHash fingerprinting and AES-based hashing.

Most C++ and Rust thread pools run 10x slower than OpenMP on fork-join workloads.

Moving beyond triple-bracket kernel launches to production-ready CUDA: proper error handling, CUDA Driver API, cooperative groups, and inline PTX for robust GPU code.

Exploring the 79-character mma.sp::ordered_metadata instruction for Tensor Cores on Hopper H100s, plus insights on PTX, SASS, and the evolving complexity of GPU ISAs.

By exploiting CPU port differences between Intel and AMD, interleaving FMA with addition instructions boosts AVX-512 reduction kernels from 211 GB/s to 330 GB/s per core.

Custom allocators can dramatically reduce JSON parsing overhead, but singleton dependencies lurk everywhere—from malloc to std::isspace—limiting multi-threaded gains.

A 16-year-old StackOverflow question on string splitting gets a modern answer: SIMD-accelerated tokenization that's 10x faster than STL and cleaner than ranges.

Nine years into a 40-year commitment to AI infrastructure.

Why SIMD programming is harder than it looks: exploring cosine similarity optimization across x86 and Arm architectures, from basic AVX2 to cutting-edge SVE2 instructions.

Achieving 5x faster set intersections on AWS Graviton 4 using Arm's SVE2 and Intel's AVX-512 with specialized instructions like HISTCNT, MATCH, and VP2INTERSECT.

Keyword arguments in Python aren't free.

NumPy's binding overhead wastes up to 90% of BLAS throughput on dot products; SimSIMD recovers that performance with optimized Python interfaces.

C++ STL strings bring 20,000+ lines of code to every compilation, yet remain slower than LibC and error-prone.

Indexing 28 billion chemical embeddings from 7 billion molecules using optimized Jaccard distance kernels, achieving 99% recall at 3,700 QPS on AWS Open Data.

Binding USearch to 10 languages reveals the friction in each ecosystem—from Python's simplicity to Java's complexity—and why C++ still rules performance-critical code.

From pure Python to AVX-512 assembly, optimizing cosine similarity reveals a 2,500x speedup through SIMD, FP16, and VNNI instructions on modern Intel CPUs.

Hand-written AVX-512FP16 SIMD code achieves 119x speedup over GCC's auto-vectorization for Jensen-Shannon divergence, using reciprocal square roots and FMA instructions.

Supercharging JavaScript's TypedArray with C bindings and SIMD achieves 10x speedup over native Arrays for AI vector operations in Node.js.

Switching from PyBind11 to direct CPython C API bindings reduced StringZilla's call latency by 5x, proving that understanding CPython internals pays dividends.

AVX-512 and Arm SVE bring masked loads and native f16 math to vector distances.

Replacing billion-entry preference lists with vector search indexes makes the Nobel Prize-winning Stable Marriage algorithm scale to real-world dating and database joins.

StringZilla uses SIMD tricks to hit 16 GB/s substring search, beating standard libraries by 5-10x and parsing multi-terabyte files Python couldn't handle.

From junior to expert: four C++ solutions to count unique strings reveal 30x performance gaps.

Vector search isn't just for AI embeddings.

A love story cut short — remembering Sona, who taught me what it means to truly care for others, build with purpose, and find someone who shares your frequency.

Write faster C++ with Google Benchmark: measure everything from math micro-ops to sorting algorithms, and discover 60x speedups with simple compiler tricks.

Why reaching DDR4's theoretical bandwidth is nearly impossible.

GPU code beats optimized CPU parallel reductions by 10x, reaching 879 GB/s.

M1 Max with DDR5 challenges 64-core server workloads in hash-table benchmarks.

DDR5, PCIe Gen5, CXL, and new CPUs from Intel, AMD, and NVIDIA converge in 2022.

Analysis of 2,000+ Linux and macOS binaries reveals less than 1% of instructions use SIMD, despite CPUs dedicating massive die area to vector processing.

A data-driven argument for why Artsakh must be independent — examining historical claims, demographic evidence, and the indisputable pattern of ethnic violence.

Seven stories tracing Turkey's path from secular reforms to neo-Ottoman ambitions — examining crimes unpunished, allies betrayed, and the dangerous pattern repeating today.

Three countries, vastly different resources and freedoms.

Why Armenia is an emerging tech hub worth your attention — from startup tax incentives to ancient history, discover what makes this small nation punch above its weight.

The pandemic revealed our weaknesses, but also unprecedented opportunities — from personal growth to financial markets, here's why optimism isn't naive.

A 2018 conversation about AI safety, infrastructure, and the real dangers of weak AI in the hands of organizations optimizing for power over people.

An iOS developer's critique of Apple's 2016 keynote — when emojis and animations overshadowed the innovations developers actually needed.