Planetside Software Forums

What should we develop on the GPU first? Let us know by submitting your vote in the poll above!

I can't vote because I can't differentiate between terrain and clouds but I think that if we could move both over to the GPU for rendering that would be a huge bonus.
Asset generation is less important than rendering in my opinion. 

We can generate both terrain and clouds quickly in Terragen already (and we can generate both in 3rd party apps/externally) and the previews and good enough to get useful results.

The slow down in creation comes from the rendering process so if GPU acceleration is on the cards that is where I would like to see it happen

With Cyphyr,  that my opinion holds much weight. A focus on a GPU renderer or hybrid would be awesome, and seems most logical route.

Completely agree with both, so I ticked clouds rendering, but terrain would be my second choice. So rendering it is.

Rendering. I ticked clouds too but hard to choose. Even just to see this poll is awesome :)

This is interesting - I'd like to see what the expected improvement in each area is.
I'd think that since GPU is heavily restricted in terms of available memory and knowing a voxel cache can take a lot of memory (for high res detailed images it can easily reach tens of GB) I wonder if GPU falls back to CPU speed once available GPU RAM is exceeded by the scene's usage.

And what's does 'Asset generation' actually mean? OpenGL performance when rebuilding a terrain or IR? Or new tools? Really can't tell so hard to take part in the poll for me.

Quote from: pokoy on June 28, 2020, 08:16:27 AMI'd think that since GPU is heavily restricted in terms of available memory and knowing a voxel cache can take a lot of memory (for high res detailed images it can easily reach tens of GB) I wonder if GPU falls back to CPU speed once available GPU RAM is exceeded by the scene's usage

I don't think this matters with GPU rendering, like a TG render is split into buckets, the workload for the GPU is also split. And in some scenarios, works a lot faster than the CPU. Hybrid renderers seem to be the best moving forward, as what the GPU can do fast, it can do, and what the CPU does best, it can do.

Quote from: pokoy on June 28, 2020, 08:16:27 AMAnd what's does 'Asset generation' actually mean? OpenGL performance when rebuilding a terrain or IR? Or new tools? Really can't tell so hard to take part in the poll for me.

I was curious about this too. Not sure what it means either. Elaboration on what these targets are for would be nice.

I'm with Cypher as well. If cpu cores could be servers to banks of gpu cores, I don't think you'd need all the data on the gpu card at once, not like you would for real time applications.

I would push all procedural stuff to the GPU, like noise algorithms etc. This should affect everything in positive way. Referring to the poll, terrain generation at first. For me, the most important thing is viewport terrain generation boost for faster feedback. Seriously. When it comes to rendering... renderer is already nicely optimized and fast, but I would opt for RTX and Optix for denoising in later turns. Thanks for asking!


Best

Quote from: amandas on June 29, 2020, 06:19:35 AMrenderer is already nicely optimized and fast, but I would opt for RTX and Optix for denoising in later turns

Totally! A GPU denoiser is the first thing I would love to see implemented.

Quote from: amandas on June 29, 2020, 06:19:35 AMWhen it comes to rendering... renderer is already nicely optimized and fast, but I would opt for RTX and Optix for denoising in later turns.

TG Path Tracing is pretty darn dog slow compared to other leaders, and with it not being NUMA aware, it inherently is not well optimized or fast. Think Cycles showing up Threadrippers 256 threads at EPYC Horizon. Finally overcoming some of Threadrippers quarks is a big stepping stone. The fastest TG render is in the 3 minute mark, which is bottom tier for most other renderers with performance on the Ryzen 5's. Lol For example, Cycles slowest 2.80 render was 427, with the very next tier being 270 seconds (Ryzen 3400G and 2600X respectively). These renders were done with 32 samples and 32x32 tiles.

Quote from: amandas on June 29, 2020, 06:19:35 AMReferring to the poll, terrain generation at first. For me, the most important thing is viewport terrain generation boost for faster feedback. Seriously. When it comes to rendering... renderer is already nicely optimized and fast, but I would opt for RTX and Optix for denoising in later turns. Thanks for asking!

I may have miss understood the question in the poll. Yes, faster feedback. Procedural work sent to the GPU.

Quote from: WAS on June 29, 2020, 12:58:55 PM

Quote from: amandas on June 29, 2020, 06:19:35 AMWhen it comes to rendering... renderer is already nicely optimized and fast, but I would opt for RTX and Optix for denoising in later turns.

TG Path Tracing is pretty darn dog slow compared to other leaders, and with it not being NUMA aware, it inherently is not well optimized or fast. Think Cycles showing up Threadrippers 256 threads at EPYC Horizon. Finally overcoming some of Threadrippers quarks is a big stepping stone. The fastest TG render is in the 3 minute mark, which is bottom tier for most other renderers with performance on the Ryzen 5's. Lol For example, Cycles slowest 2.80 render was 427, with the very next tier being 270 seconds (Ryzen 3400G and 2600X respectively). These renders were done with 32 samples and 32x32 tiles.

You mention small tiles so I assume you mean GPU rendering here?

Quote from: amandas on June 30, 2020, 06:33:36 AMYou mention small tiles so I assume you mean GPU rendering here?

Huh? No, I mean what I said... this is CPU rendering. GPU rendering is a separate category for benching cycles, and another probably for hybrid.

For me rendering happens in times I do not use my workstation or it goes to another machine over night.
There are more options to get your final render done - may cost you extra but they are out there.

And since we can only choose one option I am all in for the fastest 3D preview update we can get.
There is no substitute for that. Will RTP be developed further when the developers work on GPU processing now?
It`s a pain to wait for rerendering of the terrain. This really slows things down for me.
If the accuracy of the previz could be enhanced also - even better.
Clouds seem to be quite ok with RTP for now.

Development cycles for TG seem slow (to me as a customer), so workflow enhancement while I am building has no.1 priority.
CHeers, Klaus

Quote from: KlausK on June 30, 2020, 05:25:47 PMFor me rendering happens in times I do not use my workstation or it goes to another machine over night.
There are more options to get your final render done - may cost you extra but they are out there.

And since we can only choose one option I am all in for the fastest 3D preview update we can get.
There is no substitute for that. Will RTP be developed further when the developers work on GPU processing now?
It`s a pain to wait for rerendering of the terrain. This really slows things down for me.
If the accuracy of the previz could be enhanced also - even better.
Clouds seem to be quite ok with RTP for now.

Development cycles for TG seem slow (to me as a customer), so workflow enhancement while I am building has no.1 priority.
CHeers, Klaus

Honestly a better 3d preview would be nice, but with the current inconsistencies I always find a small full render to provide the best feedback, and can render pretty quick given the right settings. This is where the "quick renders" came from in the past when th 3D preview was even slower on old CPUs.

My vote is going definitely to rendering. This is where Terragen lags behind the most in my opinion. All major renderers are going GPU or hybrid CPU and GPU, and not only simple path tracers (like in Terragen), but also more sophisticated spectral ones with extra wavelength dimension to calculate..

The devs even have an option to adopt free of charge existing renderers like AMD Prorender, Blender Cycles or Luxcore render, instead of translating existing code to CUDA or OpenCL.

As for denoising, OIDN (Open Image DeNoise) - which is CPU based, is much better then GPU based Optix. OIDN takes maybe around 5 seconds to denoise 1080p image with 6-core CPU. Optix takes maybe 1-2 seconds on GTX1080.

You can play around with these two and compare them for free in Blender 2.90 Alpha. You can actually test it on a render from terragen.

Implementing denoiser such as OIDN in Terragen should be relatively easy, as this is or can be independent from the whole rendering process, with most wow effect. It is already available to us for free in blender for example, but it takes extra few minutes to set it up.

Anyway, I hope a lot of new potential Terragen users will see this thread, and seeing the future is bright, they will be finally tempted to make a purchase, which in turn brings the company more money, which hopefully translate to investing in more developer manpower, and by 2022/3 we might have a beta build to test and see some GPU in action.

Exciting times ahead. Can't wait.

can´t vote anymore,
But I would say rendering of cloud and atmosphere.
Rtp is decent up to a certain size of the viewport, then becomes too slow.

comparing Blender GPU and lightwave CPU when rendering vdb volumes, the cpu is just too darn slow, then again the GPU can only handle what I have in memory so to large vdb datasets, and the cuda fails to initiate.

But if that can work great with GPU for cloud rendering in the previewer/RTP   that would be it for me, I do not have the higher end cpu´s, so my mid range asus G20B rog with Nvidia gtx 1080 and 32 mb of ram and 3.2 GHz cpu will have to do...for a couple of years.

I'm sad that I am too late for voting! I would say that the preview rendering (terrain highest priority) in TG should be done first. If preview was near instantaneous, I'd be pretty happy. It is the interactive stuff that is most frustrating to have to wait for. The final rendering is generally unattended so I can sleep when that is going on.