Hacker News Reader

Comment by mourner | Hacker News Reader

Would love to learn more, since it seems counter-intuitive — are there any studies / benchmarks on this for modern hardware? Why would a uniform mesh with, say, 10x more triangles perform better if it's not updated often (e.g. a few times a second)?

View on HN

Mainly because GPU throughput is increasing much faster than CPU throughput. Spending cycles removing triangles doesn't matter if the GPU can just churn through them like nothing. In the context of a moving camera you want to reuse as much data as possible between terrain updates. With regular nested grids a ton of data can be reused, only the height needs updating as you move the grids around (and height can be reused between adjacent grids half the time if they have a power of two resolution relationship).

But baking irregular meshes into a tree-structure could be powerful too I suppose, I'm not that familiar with algorithms involving them. All of this is highly dependant on actual size of terrain and LOD requirements, terrain rendering is a deep rabbit hole.

Yeah, it's a thin balance, and to me as someone without years of experience in graphics programming, GPU feels like such a blackbox — hard to predict what will be the actual bottleneck until you implement an approach fully, with real-world loads, and even then it's hard to assess, especially with such variety of GPU performance characteristics on different devices.

So what's simpler / easier to implement in a given system will play an important role, and so are any additional things you want to do with the height data like collision detection, querying, data analysis etc. Anyway, that's a super-exciting topic to learn more about! Thanks a lot for all the thoughtful comments.

It'll come down to how frequently you're doing updates. A few times a second will certainly cost more than a prefab mesh with a texture fetch shader.

If you can precompute various LOD chunks for each tile and cache them, under some circumstances that's optimal, but I don't think any of them are webgl. A typical heightmap based terrain renderer today will consist of several (3-5) prefab LOD meshes deformed by height map textures in shader, like[1]. Swapping a textures in and out of vram is cheaper than swapping geometries, even if those geometries are of lower resolution.

Modern GPUs are almost never bound by the number of triangles on screen. The bottleneck will be the fillrate and bandwidth and pipeline stalls, especially in webgl. If you can throw down more triangles to do less of the other things, it's almost always worth it.

I have a toy mapbox tile renderer based on this approach. I'd love to know if you guys have plans to implement full 3d like google maps any time soon.

[1]: https://imgur.com/a/51g2bB4

I would really love to see your toy Mapbox renderer — please share it, even if unpolished! We'd benefit a lot from fresh ideas on how to improve Mapbox GL performance.

GPUs are effectively “magic” and can do particular things hundreds or thousands of times faster than CPUs. Offloading a per-vertex or per-pixel task to the GPU, especially if doing so reduces bandwidth requirements, is almost always a big win.

I think reality is closer to this: GPU rasterization is 20-50 times faster than CPUs and throughput computing maybe 5-20 times faster.

For example consumer Ryzen 3 with 12 cores peaks at 32 FLOPS/core/cycle. So at 3.8 GHz, peak ~1.4 SP TFLOPs (or 0.7 DP TFLOPs). But just 50-60 GB/s memory bandwidth limits it somewhat.

Quick googling says consumer Nvidia RTX 2080 peaks at 10 SP TFLOPs (or 0.314 DP TFLOPs, yes, less than half than the CPU example). Memory bandwidth being at 448 GB/s.

GPUs win massively at rasterization, because they have huge memory bandwidth, a large array of texture samplers with hardware cache locality optimizations (like HW swizzling), texture compression, specialized hardware for z-buffer tests and compression, a ton of latency hiding hardware threads, etc.

But they're definitely not thousands or even hundreds of times faster.

Reply on news.ycombinator.com