Crytek Shows Off Neon Noir, A Real-Time Ray Tracing Demo For CRYENGINE

[Stefem]

They didn't have to compare performance, they are selling the engine, not non-RTX GPUs.

I've probably misunderstood you question, didn't you asked if dedicated hardware being faster was missing in the CryTek demo?

Vulkan doesn't have anything like DXR (a very specific set of instructions to do certain thing with rays).

Yes it does, NVIDIA proposed several extension for raytracing that have since been integrated into the API and that with the some contribution of both Intel and AMD.
https://www.khronos.org/registry/vulkan/specs/1.1-extensions/html/vkspec.html#VK_NV_ray_tracing

Well, why, they said that they used SVOGI or Sparse voxel octree global illumination

That's not raytracing and it's nothing new for them nor for others as there are many games out by years that make use of voxel for GI and, BTW, it's not the solution used to render reflection (the only thing they claim is raytraced).

 

[bug]

I'm not assuming, I'm seeing it.

You're seeing what? A running demo that has no counterpart running on said dedicated hardware?

 

[medi01]

I've probably misunderstood you question, didn't you asked if dedicated hardware being faster was missing in the CryTek demo?

No. Puzzled by your interpretation.

NVIDIA proposed several extension for raytracing

Vendors are free to create extensions.

that have since been integrated into the API

Would you mind to link it? And "integrated" should not be something with vendor keyword in it.

That's not raytracing

It is ray tracing.
Voxels are much easier to check for intersections than triangles, obviously that's the trick.

games out by years that make use of voxel

Not for RT purposes.

A running demo that has no counterpart running on said dedicated hardware?

It is using different algorithm to do RT.
You cannot have "counterpart" as in direct copy of exact same demo running utilizing voxels.

Hardware accelerated 3D was hands down faster, easy to spot, one didn't need "exact same engine" to see it.
That is not the case here.

 

[Stefem]

I'm not assuming, I'm seeing it.

I'm sorry but you can't as you don't even know what and how the engine is doing, you don't know how deep the reflection goes, you don't know what kind of primitive they are using...
For clarity (even if it's well know fron day one) RT cores can't accelerate the traversal with custom primitives (works with triangles) but to be honest they don't need to, raytracing on GPUs have problem only with triangles as the horrible resulting access pattern skyrocket the cache misses. Here you can read what an engineer (a quite famous one) has done, creating a scene with a complex fractal defined structures made of 48 million spheres (perfect sphere, ther's no triangle) with shadows and reflection that goes 5 levels deep (reflectio of a reflection...) that run at 173 FPS on a RTX 2080 Ti, the demo is opensource and can be downloaded by everyone https://devblogs.nvidia.com/my-first-ray-tracing-demo/

Why would the mentioned demo, focusing on the RTX, would not be made to look good?

That's exactly what he said, compare it to the Star Wars demo (demo vs demo) and not to a game

Vendors are free to create extensions.

Would you mind to link it? And "integrated" should not be something with vendor keyword in it.

Are you trolling or have problems reading? the link was there but you deleted from the quoted text... and there you can also read the name of the AMD and Intel respective engineer.

It is ray tracing.
Voxels are much easier to check for intersections than triangles, obviously that's the trick.

Tecnically SVOGI trace cone, not rays and has its own limitation that's why it's considered a different thing from raytracing but perhaps we should first define what we mean when we use that word as the process of casting rays is used for a plethora of purpose, from physics simulation to bullet trajectory and hit detection and even for screen space effects.

 

[medi01]

I'm sorry but you can't as you don't even know what and how the engine is doing

I need to know how it is working inside, to see reflections? Good one.

the demo is opensource

That's interesting, although, not relevant.

That's exactly what he said, compare it to the Star Wars demo (demo vs demo) and not to a game

That's what I compared it too. I do not

play or even own that Star Wars game.

Are you trolling or have problems reading?

No, ARE YOU TROLLING OR HAVE PROBLEMS F*CKING READING?
Maybe it's easier if it is in red? Mkay, let me try:

And "integrated" should not be something with vendor keyword in it.

trace cone, not rays

perhaps we should first define what we mean when we use that word

It is pretty clear what *you* mean and it has a name: No True Scottsman.

 

[bug]

It is using different algorithm to do RT.
You cannot have "counterpart" as in direct copy of exact same demo running utilizing voxels.

And once again. Crytek managed to do RTRT using just general purpose hardware. We don't know how they did it, whether they hacked something or not, we don't details like poly count or the number of rays simulated. And somehow from all that you drew the conclusion that specialized is not needed.
And I'm not saying what Nvidia did was the Holy Grail of RTRT and anything different is doomed to fail. I'm just amazed at how your mind works.

 

[medi01]

@bug

details like poly coun

Poly count of what? Just how ridiculous can green BH-n get, Dear God...

Right here right now Crytek has rolled out demo with all the RT goodness that nVidia claimed needed specialized cores.
This alone is enough for RTX, that had already halved NVDA stock price, to shift to "it's dead, Jim" area.

"But muh NVDA''s stuff could be fasta!" - it sure could be, but if one can't literally see the difference with naked eyes, it's game over.

 

[Vayra86]

Sorry for digging up an old post but I think you are off base with this.
- Die space cost for RT cores is 10-15%, probably less. I am not sure if that is exactly massive.
- RTX is proprietary, DXR is not, Vulkan extensions may or may not turn out to be proprietary depending on what route the other IHVs take.
- Software-based implementation - or in this case, implementation running on general-purpose hardware - is simply not as efficient as dedicated hardware. So far everything points at this being the case here, whether you take Nvidia's inflated marketing numbers or actual tests by users. This shows even with production applications and Turing vs Titan V. RT cores simply do make a big difference in performance.
- Quality of demo is a different topic, CryTek is selling the engine so it needs to look beautiful. This one is probably best compared to the Star Wars demo. Metro is an artistic problem rather than technical one.

CryTek said this is on the release roadmap in 2019 so all the performance aspects should be testable eventually. I would expect them to talk more about it during GDC as well.

Understood and I know just as little about what RT will look like in the future. What I do know is that there are multiple ways to do it architecturally. Right now, Nvidia is already using those RT cores as an add-on to the shader itself. I'd like to see it go one step further: integration in a way that the hardware is completely programmable for many things that might be RT or might be something else. In the end that is what has happened before and given the way RT needs to be integral to almost everything that happens graphically, I think that is a sensible approach. Perhaps these cores can double up for physics calculations, for example? There are quite a few other, predictable workloads to put there, like post processing steps.

The only reason I don't believe in Turing's approach is the economy of it. This is a lot of die space to reserve and it sure is more than 15%. Because to accommodate the new featureset, L2 cache has also been expanded, for example. You can see in the relative performance to Pascal that this is also a sacrifice and a bonus depending entirely on the game in question. Turing cards hop around their Pascal equivalents everywhere by quite a margin in quite a few games. And realistically what you need to do is look at relative performance to Pascal related to die sizes. Thát is truly what you need in terms of space. Don't forget this was also a little shrink.

Let's take the 1080 vs 2060; Realistically (performance wise) we should be using a 1070ti; but let's cross that off versus the new node to get the ballpark idea.
314 vs 445mm²
29,43%; let's make it 25%.

Now count that versus what they are actually doing with this space: a few limited effects that still harm FPS.

 

[Fluffmeister]

And once again. Crytek managed to do RTRT using just general purpose hardware. We don't know how they did it, whether they hacked something or not, we don't details like poly count or the number of rays simulated. And somehow from all that you drew the conclusion that specialized is not needed.
And I'm not saying what Nvidia did was the Holy Grail of RTRT and anything different is doomed to fail. I'm just amazed at how your mind works.

Yeah it's not perfect for sure, the LOD takes a nose dive on the reflections:

So they are cheating too, hey ho, looks pretty enough anyway.

 

[bug]

Yeah it's not perfect for sure, the LOD takes a nose dive on the reflections:

So they are cheating too, hey ho, looks pretty enough anyway.

I believe Nvidia just extended DXR support to Pascal. According to them (to be taken with a grain of salt), without RT cores you get 3x less performance. It's probably what Vega has to deal with as well.

Though to be fair, I will reiterate what i said in defense of DLSS: do you really have the time to look at those details up close when playing a game? Sure we want games to look as lifelike as possible, but is it really worth it to use twice the computation power in exchange for 5 or 10% more realism?

 

[Fluffmeister]

Agreed, fact is with realistic lighting they might actually look worse. We are pissing in the rain either way, CryEngine doesn't get a lot of love from devs as it stands anyway.

 

[londiste]

And realistically what you need to do is look at relative performance to Pascal related to die sizes. Thát is truly what you need in terms of space. Don't forget this was also a little shrink.

Let's take the 1080 vs 2060; Realistically (performance wise) we should be using a 1070ti; but let's cross that off versus the new node to get the ballpark idea.
314 vs 445mm²
29,43%; let's make it 25%.

Now count that versus what they are actually doing with this space: a few limited effects that still harm FPS.

It is more than 25-30%, GP104 has 4 more SMs - 2 Pascal SMs, equal to 4 Turing SMs. However, most of it is taken up by stuff added in Volta, not RT Cores and couple other minor additions in Turing. I tried theorycrafting on this and reduced to the same SP count Volta SM is about 45% bigger than Pascal, Turing's is 10% bigger than Volta and 15% bigger that Pascal. A lot of the extra compute stuff that came in with Volta takes up a lot of space. Unsure how much of this is Tensor cores - that may be the real space hog here. I would still say RT Cores and supporting logic is 10-15% to Pascal, probably at the lower end of that scale.

 

[Super XP]

This is great.
Another Ray Tracing solution.

 

[AmioriK]

Sorry for digging up an old post but I think you are off base with this.
- Die space cost for RT cores is 10-15%, probably less. I am not sure if that is exactly massive.

the RT cores are possibly quite small, as they can be optimised EXACTLY for BVH.

Look here: https://misdake.github.io/ChipAnnotationViewer/view.html?map=TU106&commentId=470099401

Take with Grain of Salt. It is not confirmed but If that is the RT core, they are not huge at all. It is actually smaller PHYSICALLY than a single Turing SM containing 64 CUDA cores, but will most certainly delivery significantly more performance for the area used in Ray Tracing. Keep in mind that block seems to also contain other functions too.

edit screencap for convience for mobile users, chip is TU106

 

[Super XP]

Yeah it's not perfect for sure, the LOD takes a nose dive on the reflections:

So they are cheating too, hey ho, looks pretty enough anyway.

For slightly moving water, due to the Drones Fans, that is actually not bad.

 

[goodeedidid]

++++++++1 I agree 99999%! I want graphics to push hardware again and be in awe with it! I even wrote a thank you message to Crytek when I noticed they responded to one of the comments in the YouTube video. I want a new Crysis and SWEET SWEET graphics to make our machines crawl. Never before and since then has a game made me just want to look at and appreciate the visuals the same that Crysis did. Just blew me away

You know that people have to work to make those games, get salaries and holidays, teams, managers. Games don't just appear out of nowhere...

 

[KrachB00Mente]

That is all great for developers and such. But come on Crytek, give us the game that started it all and showed us what the cryengine cut really do. I am off cause talking about Crysis. I want a Crysis 4 that shows the Cryengine newest advantage with ray-tracing, DX12 and of cause the glories graphics the games are so known for and make the GPU´s sweet once again.

Please Crytek. Im tired of demos after Demos. I want crysis

Crysis cut a lot of corners like faking volumetric lightning etc. to look like it did. Thats why the performance scales really bad on modern machines.

 

[Warsaw]

You know that people have to work to make those games, get salaries and holidays, teams, managers. Games don't just appear out of nowhere...

Huh? Kind of an odd thing to comment. Never said anything about demanding a game or discussing as to why they are not if they are not financially situated. I'm just so excited for them to make another game when/if they do. That's all

 

[Deleted member 24505]

That is all great for developers and such. But come on Crytek, give us the game that started it all and showed us what the cryengine cut really do. I am off cause talking about Crysis. I want a Crysis 4 that shows the Cryengine newest advantage with ray-tracing, DX12 and of cause the glories graphics the games are so known for and make the GPU´s sweet once again.

Please Crytek. Im tired of demos after Demos. I want crysis

Actually, the game that started it all was I robot, which was the first game to use polygon graphics, which all games use now.

 

[londiste]

Cryengine has published an interview about this demo:

CRYENGINE | How we made Neon Noir - Ray Traced Reflections in CRYENGINE and more!

Ahead of GDC 2019, we revealed Neon Noir, a research and development project showcasing real-time mesh ray traced reflections and refractions created with an advanced new version of CRYENGINE’s Total Illumination real-time lighting solution. Needless to say, we did receive a lot of questions...

www.cryengine.com

 

[Vayra86]

Cryengine has published an interview about this demo:

CRYENGINE | How we made Neon Noir - Ray Traced Reflections in CRYENGINE and more!

Ahead of GDC 2019, we revealed Neon Noir, a research and development project showcasing real-time mesh ray traced reflections and refractions created with an advanced new version of CRYENGINE’s Total Illumination real-time lighting solution. Needless to say, we did receive a lot of questions...

www.cryengine.com

Very interesting read. He also touches on the bullet casings:

 

[londiste]

Short version: Lower LOD as well as cutoff from RT to Voxel reflections based on material roughness.

 

[Vayra86]

What's more interesting about that is the amount of customization they can apply to a scene to scale performance. This is a hundred times better than those rigid RTX quality levels (that need a tweak for each game as well). Especially because its integral to the rest of the scene build; there is really no hard distinction between 'RT or not'.

 

[londiste]

What's more interesting about that is the amount of customization they can apply to a scene to scale performance. This is a hundred times better than those rigid RTX quality levels (that need a tweak for each game as well). Especially because its integral to the rest of the scene build; there is really no hard distinction between 'RT or not'.

RTX does not have rigid quality levels. Indeed, Battlefield V interviews show pretty similar customization options and performance scaling choices. The guys working on Unreal Engine also went through similar process.

I have a feeling you have a bit incorrect understanding of what RTX is or does. The only part of RTX relevant here is RT Cores. What RT Cores deal with is ray tracing itself, casting rays and calculating intersections. That is it. Optimizations like preparing the structures, setting up materials/surfaces and even placing rays are generic and have little to do with RT Cores.

What CryEngine guys are showing and talking about shows the progress of effects and solutions used. CryEngine does lighting and reflections with voxels (SVOGI). Voxels have been researched and used in other methods as well like VXAO or VXGI (in this case, both from Nvidia). The way how the voxel data structures are built and handled puts them halfway towards ray-tracing in principle. Voxels are simpler to handle and less accurate but are also much faster to work with - performance hit is considerable compared to more common methods (like SSAO or HBAO in case of ambient occlusion) but compared to raytracing it is very fast.

In case of this demo, this ends up being very nice for CryEngine - their cutoff from RT is not a clearly visible cutoff but fallback to their existing voxel-based solution. It is not as accurate but when used creatively - looks like this was used for example in reflections on rougher concrete in this case - it is good enough. Both DICE and Epic have said that semi-reflective surfaces are more complex to handle with RT so not doing RT on these is a performance benefit. I wonder if this is something they can automate to a certain degree in the engine.

 

[bug]

What's more interesting about that is the amount of customization they can apply to a scene to scale performance. This is a hundred times better than those rigid RTX quality levels (that need a tweak for each game as well). Especially because its integral to the rest of the scene build; there is really no hard distinction between 'RT or not'.

Yes, what Nvidia did was showing that RTRT is borderline possible. It doesn't mean the way Nvidia did it is the best way or the only way to do RTRT. That's why I keep saying, the sooner devs stop talking thrash about RTRT and put their minds to figuring out ways to do it, the sooner we'll have nicer graphics overall.