Arseny Kapoulkine

Independent. Previously: TF@Roblox

meshoptimizer, pugixml, volk, calm, niagara, qgrep, Luau

GitHub: http://github.com/zeux
Blog: https://zeux.io

Last Notes

2026-07-28 16:57:20 UTC

- reply

units are msec, so around a second total

2026-07-28 16:30:30 UTC

- reply

Used to take ~15 seconds end to end 😅
https://cdn.masto.host/mastodongamedevplace/media_attachments/files/116/998/551/591/103/295/original/6c93f3fd649454c3.png

2026-07-28 16:22:50 UTC

- reply

Alright whatever I will just merge all textures in an array. 256 textures at 1K resolution is just ~1.4GB VRAM, and VRAM is cheap these days!

2026-07-27 21:04:24 UTC

There’s still no support for indexing arrays of textures in WebGPU right?
I found that many of my JS perf issues can be worked around by “just” running code on GPU ;) but I need access to textures and it’s difficult to do that with just uniforms…

2026-06-17 16:07:33 UTC

New blog post! In "Zigzag decoding with AVX-512", we look at two different ways to possibly (caveat emptor) speed up various forms of zigzag decoding with AVX-512 functionality that isn't just "wider registers".
Boosts appreciated!
https://zeux.io/2026/06/17/zigzag-decoding-avx512/

2026-02-18 18:15:43 UTC

I don’t fully understand the motivation to use LLMs to write blog posts. The loss of personal voice is so striking that it immediately rings alarm bells in my head. You can fact check the output although even that is spotty but you can’t inject the soul back.
And if you stop writing yourself, what makes you think you still can? And what is it that makes you - you?

2026-01-06 03:37:16 UTC

- reply

Congrats!

2025-12-10 19:23:00 UTC

fun
https://cdn.masto.host/mastodongamedevplace/media_attachments/files/115/696/898/857/775/075/original/0fb95098ab0d4dd7.jpeg

2025-10-01 18:06:33 UTC

- reply

You can set the error to be higher in theory. I think it depends on the particulars of your TAA; and also because actual Nanite does micropoly rasterization in compute for small triangles, I would imagine the performance depends on rendered triangle count in a non-trivial way. Note that unlike discrete LOD, under camera motion you’d expect many cluster transitions every frame in different parts of the screen… so single pixel cutoff is much safer.

2025-09-30 17:39:47 UTC

New blog post! In "Billions of triangles in minutes" we'll walk through hierarchical cluster level of detail generation of, well, billions of triangles in minutes. Retoots welcome!
https://zeux.io/2025/09/30/billions-of-triangles-in-minutes/
https://cdn.masto.host/mastodongamedevplace/media_attachments/files/115/294/467/861/286/476/original/44ba883e92d0c392.jpg

2025-04-25 20:12:51 UTC

- reply

Classic, alongside "we support vertex texture fetch but the list of formats is empty", and I think was another one like this that I'm blanking on...

2025-04-01 04:49:22 UTC

New blog post! "Measuring acceleration structures", in which we will compare BVH costs on various GPU architectures and drivers and attempt to understand the details enough on AMD hardware to make sense of the numbers!
Boosts appreciated :)
https://zeux.io/2025/03/31/measuring-acceleration-structures/

2025-03-14 16:04:08 UTC

meshoptimizer 0.23 is out! Featuring vertex codec v1 (increased compression ratio, faster decompression), even faster decompression on Apple Silicon, improvements to meshlet builder with fewer disconnected clusters and more options for clustered raytracing and hierarchical clusterization, new algorithms for hierarchical clusterization, simplification quality improvements and more.
Thanks to Valve for sponsoring most of the work on the core library in this release!

2024-08-14 16:56:36 UTC

Apparently the current state of the art in developer tooling is write it in Rust (cool), render it with Vulkan (ok) and write a custom UI framework (oh) that no other tool uses.