Vulkanification

Vulkan

Many years ago, I saw some Mantle API-related news and started thinking that it looked like something "better engineered". I was 12 at the time.

Later those ideas became Vulkan, and today everyone knows how difficult it is compared with other graphics APIs.

Does difficult mean good? Not necessarily. But with Vulkan, it means you can do low-level operations on your GPU.

GPU programming is notoriously less straightforward and less forgiving than CPU programming. That's the most important value of Vulkan — the tools to make your GPU work the way you want.

What does it mean?

I started learning Vulkan. The classic reason: it's interesting. It can be used for UI frameworks, 3D graphics and, of course, compute.

For example, Godot's renderer is built with Vulkan. At the same time, id Software used it extensively, and its games run pretty smoothly. llama.cpp has a Vulkan backend, which I use instead of ROCm.

There are many examples showing how well it works in various use cases.

But why not wGPU?

It's also great and can use Vulkan as a backend on Linux. Its main advantage is cross-platform compatibility, as well as the fact that it works on the web.

Those are magnificent qualities, but they aren't the most important ones for what I do. I want to experiment, try modern approaches, maybe even something alternative.

What's the project?

It's a classic Vulkan project — triangle. In order to show a single triangle on your screen, you touch a lot of internal Vulkan details, and there are multiple.

Vulkan is really soaked with idea of structures for every possible entity. Obviously, it makes sense because there are reasons.

ChatGPT 5.6

Well, I am a Sol Max connoisseur; it might be slow, but it finds potential issues. Luna is also the way to go, but Terra seems to be kind of useless. I tried Ultra once, and it ran for a really long time without finishing. Maybe I will try it later.

Open models?

It seems like there are many interesting releases that will be evaluated only after some time. For now, I stick to Gemma 4 QAT models of different sizes.

My friend implemented some an extension for the Hermes agent for protection against prompt injections. Overall, this CaMel approach is new and not tested.

At the same time, an idea of having one more model for stuff like intent definition or something similar is what I thought about before.

It requires more testing, but having a smaller, probably CPU-resident model might be a really beneficial idea. It really comes down to how well it can do what it's supposed to do. Maybe it's a question of fine-tuning.

Indie gem I missed

For some unknown reason, I didn't play Hollow Knight. I had heard of it and knew that people liked it. Silksong also brought it some additional popularity. Anyway, two weeks ago I bought it.

I'm currently trying to 112% the game. Sometimes the controls are extremely janky, and I have to use the keyboard. Also, I changed the controller layout because the default is questionable (Did those people have alien fingers?).

There is a lot of stuff to learn from this game, as well as content to enjoy. Absolutely great experience.