Ash Vardanian

My friend @vinogradovk is launching something new — an Open Source Endowment (OSE), a US-based nonprofit that will fund open-source work that silently powers the world. I've joined in. Maybe you should, too 🤷🏻 Over the last 12 months, I have reached out to several tech companies…

Busy-waiting is tricky! The standard approach is to spin on your standard library’s “yield”, be it in C++ or Rust, but it’s clearly suboptimal I’ll experiment & publish more on TPAUSE- & WFET-based incremental backoffs in the next months, but it may require more hardware

Given how little content survived from 5,000 years ago, it would be ironic if none of Mesopotamian “kings” were actually rulers - just literate sociopaths with a flair for pretend-governance who happened to preserve their ramblings better than others 😅 worldhistory.substack.com/p/what-were-th…

First "agentic coding" win for me 🥳 Asked Claude to rewrite StringZilla’s C → CPython glue using fast-call like SimSIMD ~4K lines of C99 boilerplate — never again by hand If Claude can do it, you can too — drop PyBind11 in latency-critical paths x.com/ashvardanian/s…

Every algorithmic feed - same pattern: 🥇 Summarizing & spreading news 🥈 Endorsing someone else's launch 🥉 Original deep-tech work Any tricks to reverse it!? Tried boosting posts on X & LinkedIn - zero impact, wasted money & effort

Half the girls’ dating profiles these days: “Looking for a high EQ guy who goes to therapy”… Not one is looking for a man with close friends and healthy relationships 😵‍💫

If memory serves me well, a few years ago I was presenting @Unum_cloud at the same @UBS-organized event where Builder.ai was participating. It was a well-run event with a handful of founders and a sea of investors — SoftBank, Insight Partners, the usual suspects. I…

Cross-generation hardware benchmarks are rare @AMD's Zen 1 to 5, 2017 to 2025, resulted in a 7-8x geo-mean improvement, with greater gains in mixed-precision numerics & AVX-512 Still waiting for Turin access on @AWSCloud to extend/test SimSIMD sparse linear algebra kernels

Tried signing up for @IBMcloud to test their hardware and maybe port some of my open-source software to their mainframes… on my own dime 2 hours later: still stuck fighting their registration form and support 🤦‍♂️

Curious if anyone’s using the 8–16 socket IBM E1080 systems for large-scale data processing outside of SAP HANA. POWER10 offers ~400 GB/s memory bandwidth per CPU, so a full system pushes ~6 TB/s. Feels like HANA doesn’t really scale that far — what does?

Wow — DeepSeek is now considered a “Frontier Model”… and apparently the best at generating C++ code Source: arxiv.org/abs/2506.17298

For Rust users, SimSIMD v6.5 now exposes mixed-precision conversion & arithmetic utilities It's time to make `f16` and `bf16` first class citizens of every data-processing pipeline 🦀 github.com/ashvardanian/S…

Just released new parallel primitives for SMT: NUMA-aware, Huge Pages friendly, and using hardware-specific busy-waiting (like TPAUSE, WFET). Dynamic dispatch included 🤗 Needs tuning, but the core is solid – Fork Union v2 is out: github.com/ashvardanian/f…