3 ms·
Whether this is a win is an difficult question. Rend3 recently put in meshlets, and performance got worse for my use case. It's a way to reduce draw work at som
by Animats 4y ago
Whether this is a win is an difficult question. Rend3 recently put in meshlets, and performance got worse for my use case. It's a way to reduce draw work at some cost in GPU memory. But in a metaverse application, where you have user-created content, you have less instancing, and tend to run out of GPU memory before running out of rendering power.
Like mipmapping, trying to do level of detail at the GPU level only results in less rendering load, not less GPU memory consumption.
There's the question of whether you want to do occlusion culling down at that level. Yes, it cuts GPU load way down when you're in a windowless room. Then go outside and the load goes way up. If you have windows in your buildings, the occlusion culling doesn't help as much. This trick improves average GPU load for a scene, but not worst-case GPU load, which is what matters for playability. When I tried this on an older GPU with slow compute shaders, the frame rate got worse by a factor of 3, and the GPU memory consumption increased.
If you want to do occlusion culling, there's an argument of doing it by pre-computing occluded zones for a level. Unreal Engine does that sort of thing. Analyze once at build time, not once per frame.
That's a problem with pushing work down to the GPU level. You have to do it on every frame, which may be a lose.
- Pulcinella 4y agoLike mipmapping, trying to do level of detail at the GPU level only results in less rendering load, not less GPU memory consumption. Though on Apple’s ARM devices (and consoles) memory is shared between the CPU and GPU. I am unsure if making some asset available to the GPU to render (e.g. a texture) requires making an extra copy in ram just for the GPU, but if it is not required and you can just have one copy, then shouldn’t doing LODs GPU vs CPU require roughly the same amount of memory?
- Animats 4y agoIf you don't need a higher LOD for a texture yet, it doesn't need to be in memory at all. It can be out in a file or cache. Big-world networked games have something that looks like a cache hierarchy. There's content in the GPU, content in main memory, decompressed content in a cache on SSD, content that hasn't been decompressed yet, and content that's still on a server somewhere. Much of this data wrangling is done at a slower time scale than the frame rate. So you don't want to push all that down to the GPU level. The heavy machinery behind this is much of building a metaverse is really about. If you don't do all this, the result is low-rez, like Decentraland and Horizon. If you do it right, things look pretty good.