Are clouds the best ever now? I'm impressed.

[Noel 4,487]

Here is a fascinating snippet from an article about Voxels and Performance.  My highlight in red makes me think very small changes in voxel size have big performance implications.  So if a slider were implemented it would be covering a very small range:

Unfortunately, the peak performance of voxels is suboptimal because the number of voxels scales as N³, and GPUs aren’t ideally suited for the update techniques we have to use to keep performance manageable. Given enough research into GPU computing we should be able to offset the performance loss, but the current baseline cost can be pretty high.

 

 

[Noel 4,487]

2 hours ago, Cpt_Piett said:

I very rarely notice the pixelation of the clouds.

Me too Cpt--but they are only a tipoff for the low-resolution, or low voxel density, of how clouds are being rendered at the moment, and I think that has major implications for how far you can go with cloud depiction quality, without addressing this low voxel density.  As an analogy, you're already zoomed in on a taxiway sign as you look out the side window of your plane.  You see the object is very close to the window--maybe 30 feet or so, just like that close-range cloud you see.  Now imagine if that taxiway displayed this low of a resolution where you could easily see the voxel grid studs--it would be awful, or circa 1990.  I believe this is what we're seeing w/ cloud rendering up close and this is why my contention is this plays a big roll in depiction quality.  Off in the distance it doesn't matter nearly as much, but plays some role.  Perhaps Asobo will use an algorithm that says the closer one gets to the cloud the higher the voxel grid density becomes.  In fact, it's quite possible this is already happening, but will benefit from tuning to let those w/ the processing horsepower take advantage of it.  I'm guessing it played a major role in the downturn of close-range cloud depiction happening sometime after SU3 or SU5.  

Edited July 31, 2022 by Noel

[March Hare 720]

Yes, voxels are costly and so is the ray marching applied to them in the cloud formations, on top of the noise algorithms (likely Perlin or perhaps simplex noise, and other types) used to "sculpt" the clouds.

[March Hare 720]

There's a reason why the voxels show up more sometimes than other times, under varying lighting conditions, and that's down to the ray marching, which by its necessary nature samples each individual voxel as it passes through and shades them individually, to the best of my knowledge.

[Rocky_53 1,566]

On 6/21/2022 at 6:53 AM, MarcG said:

Because they've more than likely got a license agreement with Meteoblue, in fact I'd put money on this being the stumbling block and refusal of Asobo/MS to fully open up the WC (which is fair enough, it's their game and their license/business deals).

Did you call it a game....? 

[asanosho 55]

16 hours ago, Noel said:

Here is a fascinating snippet from an article about Voxels and Performance.  My highlight in red makes me think very small changes in voxel size have big performance implications.  So if a slider were implemented it would be covering a very small range:  . . .

Indeed!

And the range of this slider would have to be so small that you would not be able to see any difference in cloud depiction.

To be able to see a marked difference you will have to increase the voxel density by at least a factor of 4. That would give you 64 times more voxels to render, effectively increasing the time to render a cloud by a factor of 64.

You would have to adjust your T-LOD to around 1 or 2 instead of 200 and even then your system would stall.

[SierraDelta 1,337]

20 hours ago, mrueedi said:

But not on anyones system. Mine don't look like that. What setting do you have for the volumetric clouds in the graphics setting?

Just bumping this question so that @Noel can answer. And to reiterate - some of us don't see clouds pixelated when up close.

[Keto Ketchup 240]

1 hour ago, SierraDelta said:

And to reiterate - some of us don't see clouds pixelated when up close.

I do, with ultra-quality clouds, so I'm interested to learn how some of you are avoiding this.

[Noel 4,487]

4 hours ago, asanosho said:

Indeed!

And the range of this slider would have to be so small that you would not be able to see any difference in cloud depiction.

To be able to see a marked difference you will have to increase the voxel density by at least a factor of 4. That would give you 64 times more voxels to render, effectively increasing the time to render a cloud by a factor of 64.

You would have to adjust your T-LOD to around 1 or 2 instead of 200 and even then your system would stall.

I've think you've oversimplified the story greatly to make a point but I think the general idea is probably in the right direction, impacted also by DX12 and driver optimizations for these voxel matrices.  What does a factor 4 do to voxel density?  Why 4 and not 2?  If I start with a density of 100 voxels per cubic area X, what factor are we talking about to go to 110 voxels per same cubic area?  Are voxels quantized such that they can only be changed in discreet amounts?

When MSFS first launched cloud density was superior to what it is today such that when you got close to clouds you did not see this degradation we have today, post SU3/5.  I was able to run the sim quite well, with clouds set to Ultra, and at that time had an RTX 2070 Super in the box which is quite inferior to what I'm running today.  The only time I had to dial back clouds to High was on occasion very diffuse, thin, fog-like clouds would increase GPU utilization.  

I don't see the evidence of the architecture (the voxel grids) all the time up close to clouds and they aren't problem per se--they just point to the source of this low- resolution rendering.

Edited August 1, 2022 by Noel

[Cpt_Piett 4,538]

1 hour ago, Keto Ketchup said:

I do, with ultra-quality clouds, so I'm interested to learn how some of you are avoiding this.

AFAIK it's not possible at the moment. Even with 4K TAA ultra clouds, there will still be some pixelation. It's the way the clouds are rendered. It can be remedied somewhat I've found by disabling film grain and sharpness in usercfg.opt, and instead tweaking these values using Nvidia freestyle filters. But still, if you zoom in close enough, you'll see some pixelation. 

[March Hare 720]

I can't be entirely sure about this, because of the nature of memory, but I'm fairly sure the "graininess" has always been present (at times). It's just logical that it has, knowing a fair amount about how volumetric cloud rendering generally works. But it's also true that we wouldn't always see it in every flight. Indeed, sometimes I do but often I don't, and it's always been that way since launch to my naturally imperfect recollection.

I have a few theories on this, one of which I mentioned before about the way ray marching works, unavoidably showing up the edges of voxels in certain lighting conditions.

The other is that some of the graininess may be some types of noise becoming visible, for example when antialiasing isn't working to its most ideal extent. Blue noise would start to show, for example, with less than ideal antialiasing, and that has been used in volumetric cloud rendering to achieve certain types of cloud detail and colouration. It may be that the Perlin or simplex noise (which is, in a way, grid based) chips away at the voxels to form the cloud shapes in such a way that can't be smoothed out.

Lastly, there could be something to the "resolution" notion: they may be using dynamic scalability, where chunkier voxels, or less voxels per cloud, are used depending on how much is happening at any given time -- more clouds, more voluminous clouds, more other rendering (denser terrain assets), and other underlying computations happening, so it intelligently adjusts to save resources.

Or a combination of all these things.

So, those who never see it, or not always (like me), would just be getting lucky.

 

Edited August 1, 2022 by March Hare

[Cpt_Piett 4,538]

The graininess is definitely infrequent. It seems to be related to lighting conditions, among others. I'm not that well versed in the technical bits of how clouds are actually rendered. But it seems to me that, every now and then, some clouds appear much more grainy than others. 

[Noel 4,487]

10 minutes ago, March Hare said:

Lastly, there could be something to the "resolution" notion: they may be using dynamic scalability, where chunkier voxels, or less voxels per cloud, are used depending on how much is happening at any given time -- more clouds, more voluminous clouds, more other rendering (denser terrain assets), and other underlying computations happening, so it intelligently adjusts to save resources.

And/or...

22 hours ago, Noel said:

Perhaps Asobo will use an algorithm that says the closer one gets to the cloud the higher the voxel grid density becomes.  In fact, it's quite possible this is already happening

I have always wondered how far this rendering method could go with respect to the kind of detail we see in clouds in the real world.  My sense is without getting much tighter voxel density it's just not going to be able to happen all that well, and that may well be out of the question for the foreseeable future.  But new technologies and approaches to 3D modeling will evolve.  Still--what I'm wondering is if one can increase voxel density by any value.  I have to think you can though the geometries involved may complicate this.  But again, I've got plenty of GPU resources to exploit and I hope they can continue to scale up where the hardware can handle it, and there's more on the way.  I can't believe my CPU's single-thread performance is already something like 28% slower in PassMark than I9-12900K.

[Cpt_Piett 4,538]

2 minutes ago, Noel said:

And/or...

I have always wondered how far this rendering method could go with respect to the kind of detail we see in clouds in the real world.  My sense is without getting much tighter voxel density it's just not going to be able to happen all that well, and that may well be out of the question for the foreseeable future.  But new technologies and approaches to 3D modeling will evolve.  Still--what I'm wondering is if one can increase voxel density by any value.  I have to think you can though the geometries involved may complicate this.  But again, I've got plenty of GPU resources to exploit and I hope they can continue to scale up where the hardware can handle it, and there's more on the way.  I can't believe my CPU's single-thread performance is already something like 28% slower in PassMark than I9-12900K.

I'm sure they will scale up in response to further developments in hardware. The 12th gen Alder Lake CPUs have AFAIK been a success for Intel, and I'm sure AMD is doing their best not to fall behind in the race. And there's also the upcoming Nvidia 4 series.

[Krakin 5,739]

People keep talking about downgrades now being present but they hardly ever show pics of when things were different