Current Generation Games Analysis Technical Discussion [2023] [XBSX|S, PS5, PC]

[iroboto]

This can't be the reason everytime someone crticizes an aspect of Starfield, he is shut down because of the 10000 object thing. No, when the game is running this suboptimally it is not rendering thousands of objects on screen, it is running regular, plain scenes found in dozens of games.


What is Starfield exactly? What does it do so differently that elevates it above the mentioned examples and justify it's significantly higher cost?

Because it does indirect lighting real time?
Why can’t it be the reason? Scenes are not levels built in a 3D modelling program to be imported into an editor in which everything is stapled to the floor.

Levels in Starfield are made by putting objects together into a setting. Chairs, desk items, tables, many of these items are movable. It’s not about 10,000 objects in a single space but the many hundreds of objects placed around that to build up a scene that can be interacted with, moved, taken and destroyed.

 

[Silent_Buddha]

Basically, for a BVH a given scene with mostly static geometry has all the geometry in predictable places from frame to frame.

In Starfield with all the interactable objects, items may or may not be in predictable places from frame to frame. Revisiting any location means the scene may not be the same as when you first visited the scene. IE - there's no way to cache the BVH or even necessarily build it ahead of time. It's not even going to have a predictable amount of interactable objects as players can remove or place objects in the 100's or 1000's. It'd have to be rebuilt everytime any object comes into view because it may not be in the same location. NPCs might bump into it, players may bump into it.

If you go with the assumption that you'll likely have to rebuild the BVH everytime a location is visited, then pre-building/caching the BVH for a location ahead of time isn't really worth it. Now throw in procedurally generated scenes with hundreds and hundreds of interactable objects.

Even RDR2 wasn't doing things on the scale of Starfield WRT interactable objects that may or may not be in a given location anytime a person visits said location. Metro: Exodus is extremely simplistic and more importantly with a much greater reliance on predictably (as in every time a location is visited) placed objects WRT scene complexity compared to Starfield as is Cyberpunk 2077 and Days Gone.

Regards,
SB

 

[troyan]

And nothing happens. Compare the "interactivity" to Control and Starfield looks just outdated.
And here a few screenshots from outdoors:

https://www.reddit.com/r/nvidia/comments/16m0f09

There is nothing to compute in these screens which justifies the performance on nVidia GPUs.

Basically, for a BVH a given scene with mostly static geometry has all the geometry in predictable places from frame to frame.

In Starfield with all the interactable objects, items may or may not be in predictable places from frame to frame. Revisiting any location means the scene may not be the same as when you first visited the scene. IE - there's no way to cache the BVH or even necessarily build it ahead of time. It's not even going to have a predictable amount of interactable objects as players can remove or place objects in the 100's or 1000's. It'd have to be rebuilt everytime any object comes into view because it may not be in the same location. NPCs might bump into it, players may bump into it.

If you go with the assumption that you'll likely have to rebuild the BVH everytime a location is visited, then pre-building/caching the BVH for a location ahead of time isn't really worth it. Now throw in procedurally generated scenes with hundreds and hundreds of interactable objects.

Even RDR2 wasn't doing things on the scale of Starfield WRT interactable objects that may or may not be in a given location anytime a person visits said location. Metro: Exodus is extremely simplistic and more importantly with a much greater reliance on predictably (as in every time a location is visited) placed objects WRT scene complexity compared to Starfield as is Cyberpunk 2077 and Days Gone.

Regards,
SB

And in Cyberpunk there are not hundreds of objects on the screen which are moving in each direction. Or in Control there are no debris and objects flying around in a fight...
You do not build the whole BVH, it gets refitted with the new objects. That was the first thing nVidia talked about five years ago.

The great thing about DX12 is that you can build the BVH with multithreading on a CPU and let the the GPU part running in async compute on the GPU. When i think about it thats sound like some people have thought about the whole process...

 

[Scott_Arm]

Good thread on why some of the outdoor scenes in Starfield can look quite bad.

https://twitter.com/x/status/1703671903431360725

 

[DavidGraham]

Because it does indirect lighting real time?

Where is that exactly? I can't see any example of that, half the biomes in the game look last so last gen it's actually embarrassing. See the post above.

And see more here.

https://twitter.com/x/status/1703711781376663589

It’s not about 10,000 objects in a single space but the many hundreds of objects placed around that to build up a scene that can be interacted with, moved, taken and destroyed.

None of that is new, we've had that since Oblivion.

None of these videos show indirect lighting affecting the objects, all I see are objects lit by direct light.

The game seems to be missing key lighting features.

https://twitter.com/x/status/1703675446624149687

 

[iroboto]

Where is that exactly? I can't see any example of that, half the biomes in the game look last so last gen it's actually embarrassing. See the post above.

It looks that way because the under shade is entirely done through indirect lighting, as per the tweet. And the lack of normals in that area combined with missing frequency on AO is what causes the flat looking. That clearly isn't direct lighting. And it's clearly not baked indirect lighting, because as cited earlier it's impossible to bake the indirect lighting for 1000 procedural worlds and 100 hand crafted worlds. Unlike all the other GI solutions you've brought up.

But if that isn't obvious enough then here is a DF video covering it:

What I find disparaging is this need to uphold some graphical standard of performance based on last generation techniques and then be upset that next generation titles don't meet that expectation. What I also find disparaging is that you want all titles to look and render the same? Should we not have engines that support different features to have support the various ways we can interact with games? Every game should not be reduced to just shooting, racing, or walking around looking at environment details.

Is this not a thread to discuss the _technology_ behind these engines, and how they work? Instead, in this thread, and many others, no discussion on pros and cons, no discussion on how this technology stack supports different gameplay experiences - absolutely no interest in the discussion of how well they were able to scale a solution all the way down to Series S. All I'm reading from the same group of individuals is some very select cherry picking a handful of environments that makes up less than 1% of the total time played or is available to see is not aesthetically pleasing. It's sad really.

The game is far from perfect, and no on disagrees that performance could be better (and ideally IHVs release new drivers to further improve on this) but the level of gas lighting of recent releases with modern rendering techniques (Real time GI, AO, and lighting) by the same people has hit new levels on this forum. Ray Tracing went through this exact same trend, when people couldn't see the 'difference', it wasn't worth the performance trade off. Now it's gospel. Well, there's no difference here for me, whether it's software or hardware based, this is where we are headed - away from baked.

This the cost, and even if it doesn't look like it, those calculations still need to take place; please get over it.

 

[DavidGraham]

But if that isn't obvious enough then here is a DF video covering it:

DF highlights the same problems talked about in the argument. Lighting is inconsistent. And even with it's inconsistency and low quality, the game remains slow to render comparative to many other games. The best DF has had to say about the indirect lighting solution is "reasonable".

Is this not a thread to discuss the _technology_ behind these engines

Before we do that we first need to identify the problem, hand waving it and pretending all is fine and dandy is the wrong attitude of discussion. Clearly the lighting of the game is in need of major improvements, many players have demonstrated that, and several devs have weighed in on the matter.

Ray Tracing went through this exact same trend

Of course this is not the same, with ray tracing there were clear differences, people just argued if they justified the performance cost. Here we neither have differences or performance. The GI system is equal or worse than other comparable systems in other games, yet it performs considerably worse.

The game is far from perfect, and no on disagrees that performance could be better

On that we agree.

no discussion on pros and cons

We are doing that indeed, technically discussing the cons of the lighting system in both quality and performance, quoting devs and players alike. While also acknowledging that it indeed looks good in interior situations.

 

[Scott_Arm]

@iroboto My interpretation of Sebastian Aaltonen's tweet thread is basically the same issue Halo Infinite has. They use a very simplistic GI solution, which is probably light probes. Once the sun (point light) is occluded you end up with a shadowed area, and anything is getting it's indirect lighting from light probes. Those light probes lack strong direction because the probes are sampled and interpolated, so light seems to come from all directions. This basically reduces the contribution of normal maps to something like zero. In terms of GI, what Starfield is doing is probably VERY simple because it seems to have the same issues as much older games that used light probes to approximate GI.

I'm guessing the daylight scenes that look flat are something where the GI doesn't have enough resolution to have good indirect shadows so everything gets a flat indirect light from all directions. I don't know.

 

[see colon]

What I find disparaging is this need to uphold some graphical standard of performance based on last generation techniques and then be upset that next generation titles don't meet that expectation. What I also find disparaging is that you want all titles to look and render the same? Should we not have engines that support different features to have support the various ways we can interact with games? Every game should not be reduced to just shooting, racing, or walking around looking at environment details.

Yes. I don't know what is going on right now where people are upset about a game, with an in house engine, using in house solutions, is being disparaged because it doesn't look like or perform like competing solutions, while at the same time complaining about consolidation in the middleware market is causing every game to look and perform the same. So yeah, Starfield's GI doesn't look as good as Lumen, or RT solutions. But isn't that what makes Lumen and the hardware RT solutions special? I'd also like to point out that Starfield's performance isn't out of line with recent Unreal Engine releases. Immortals of Aveum, for example, is often sub 60fps at 1440p on an RTX3080, and can't maintain 60FPS without upscaling. A 3080 is roughly the same in Starfield. It needs upscaling to maintain 60fps.

Parts of Starfield look good. Other parts of Starfield look good for a Bethesda game. This game uses reasonable amounts of VRAM, while maintaining high quality materials throughout, unlike several other recent releases. And it doesn't have shader stutter, unlike several other recent releases. And it launched in a fairly stable state on all platforms. Sure, there were missing features (DLSS/XESS, HDR settings, FOV, AF), though some of those might not have been added for contractual reasons, others (like AF) might be missing for technical reasons.

 

[iroboto]

@iroboto My interpretation of Sebastian Aaltonen's tweet thread is basically the same issue Halo Infinite has. They use a very simplistic GI solution, which is probably light probes. Once the sun (point light) is occluded you end up with a shadowed area, and anything is getting it's indirect lighting from light probes. Those light probes lack strong direction because the probes are sampled and interpolated, so light seems to come from all directions. This basically reduces the contribution of normal maps to something like zero. In terms of GI, what Starfield is doing is probably VERY simple because it seems to have the same issues as much older games that used light probes to approximate GI.

I'm guessing the daylight scenes that look flat are something where the GI doesn't have enough resolution to have good indirect shadows so everything gets a flat indirect light from all directions. I don't know.

Agreed, but disagree on them using light probes. It’s something else that is simplistic but Playing enough of the game, I feel its accuracy is significantly better than light probes.

 

[Scott_Arm]

Agreed, but disagree on them using light probes. It’s something else that is simplistic but Playing enough of the game, I feel its accuracy is significantly better than light probes.

I kind of wonder if they're doing something different indoors vs outdoors. Indoor areas look good. Outdoors is where it's very hit and miss and can look quite poor in the wrong conditions.

 

[iroboto]

I kind of wonder if they're doing something different indoors vs outdoors. Indoor areas look good. Outdoors is where it's very hit and miss and can look quite poor in the wrong conditions.

yes. But also it might be a physics issue. The planets are as far away as they say they are, as well as any sun from the planet. Everything moves in the constellation. So the light sources are technically super far away outdoors. Whereas indoors they are metres away. So I’m not sure if this is a reason, but it could be poor calculation on far distance light sources.

 

[troyan]

I think indoors has no light change, so they can bake most information into the texture and have only on direct light source. That is much harder outdoors with the day/night cycle.

 

[Scott_Arm]

I think indoors has no light change, so they can bake most information into the texture and have only on direct light source. That is much hardware outdoors with the day/night cycle.

I doubt they're baking lighting into textures. That's a very old way of doing things.

 

[DavidGraham]

The director of DLSS at NVIDIA believes that Starfield isn't a terribly well optimized game. (timestamped)

 

[Scott_Arm]

The director of DLSS at NVIDIA believes that Starfield isn't a terribly well optimized game. (timestamped)

I think he’s just talking shit because of the amd FSR thing

 

[techuse]

I interpreted it as being attributed to Nvidia's severe performance deficit relative to AMD.

 

[Scott_Arm]

I interpreted it as being attributed to Nvidia's severe performance deficit relative to AMD.

Probably that too.

 

[arandomguy]

I haven't watched it but he might have an idea of how much access/input Nvidia had with to Starfield relative to other titles which would drive his opinion as well.

But to me again this comes down to how we interpret the word optimization. Let's be honest here, Nvidia on the PC was likely no the optimization priority for Starfield, we don't need any deep level access to code to know that as the best optimization for Nvidia's users on the PC is DLSS over other options. Does that mean the game overall is unoptimized? Well people would say it's at least more optimized for AMD users on the PC. And more so overall it does seem optimized very well for the Xbox Series S.

That's just my view though, but I've never been the camp that there is anything such as truly neutral software, the industry is what it is.

 

[arandomguy]

Starfield is just one data point of many. The relationship between register usage, occupancy, latency and bandwidth has been discussed (mostly by Nvidia) since CUDA 1.0. I think we can objectively say RDNA 3 is better equipped than Ada to utilize available compute and bandwidth in complex workloads. Less register intensive workloads that fit in L2 will favor Ada.

But where I'm going with that comment for instance is can we assume that Starfield is a "complex workload" or that it is the best approach given the results (at least for all hardware)?

At least my impression from some of the wider reporting and narratives is there is interpretation being used here to gauge/judge the overall validity of each hardware approach from a generic stand point, when it seems at least to me subject to the software involved.

LoL, expecting game devs to maintain separate render paths for each IHV, thats seems like asking too much.
I suspect most devs maintain 1 path, ie. optimize mostly for the common case, eg NVIDIA.
and let the others be.

This isn't just about slightly different compile flags when compiling your shaders - Though that might help a bit - This is about the overall rendering design and shader complexity.
Sure CPU compilers have an favor speed/favor size flag, but i'm not sure how relevant or comparable that would be to a GPU shader.

More asking questions but isn't this one of the issues with DX12 with respect the realities of the PC market? As we are discussing shaders here specifically I thought an issue with DX12 is that shader replacement is not as straight forward of an approach from the driver/IHV side to optimize unilaterally. This essentially shifts more onus on the game developer to optimize for the wider array of architectures.