GPU Ray Tracing Performance Comparisons [2021] *spawn*

Far Cry 6 is already falling 10-20% below 60fps on RTX 3090 at 4K, so it seems like there's no room to add quality to reflections. The shadows seem fine as they're substantially better than non-ray-traced.

This gets back to the problem of devs pushing quality in a game into the "far-beyond playable" zone. If RTX 3090 were to run 4K at 30fps 1% lows in this game, it's going to cause a more substantial backlash than the half a dozen people on this thread who think the reflections should include an option for 1:1...

Devs have learnt that unplayable max settings hurt sales of their games.

Sure, this game seems to be labouring with an engine that should have been ditched - the stutters caused by driving quickly are clear evidence it's fundamentally broken. But they've chosen to limit max quality to a playable-60fps with VRR on a 3090 at 4K, so it is what it is.

 

This isn't an honest discussing about eh benefits of tessellation.
Crysis 2 was one of the most infamous "the way it meant to be played" games, "optimized" to Nvidia. Tessellation on AMD was perfectly fine but Nvidia had overdeveloped tessellation on their cards to much that they saw an opportunity. Crysis 2 used huge amount of tessellation on unsuspecting perfectly flat objects that didn't needed more than a few dozens polygons and rendered an invisible tessellated ocean below the floor.

I'm really curious, where these things removed or they stayed there?

 

As an aside, I think what I like about sampling resolution as a general problem is that it degrades gracefully. Sure, you have to deal with aliasing from undersampling but we seem to be getting better at reconstruction and antialiasing heuristics. So I would rather have an accurate* but undersampled simulation that is then heuristically reconstructed (or at least smeared to remove high-frequency artifacts).

* For sufficiently vague definitions of "accurate", so that I can weasel my way out of doing infinite**-bounce GI.
** For sufficiently low values of "infinite", e.g., 2.

Yeah I would love to see that too. I was really heartened to see ME:EE working on XSS. Incidentally my point about better simulation at low resolution is nicely captured by Alex's XOX-XSS-XSX comparison here:
I would take the XSS version over the XOX version, at least on a 1080p display. It's much tougher to swallow on a 4K display but a more sophisticated reconstruction may help.

Could you elaborate? I can think of 2 challenges with sophisticated materials:
(1) Bounce angle computation is more involved.
(2) Shader divergence problems are exacerbated with sophisticated materials (simply because the shader code is more branchy).
Anything else?

 

We had no such review guidelines.

They never existed in the first place - we have talked about this before in a video and have had multiple former crytek employees comment on it. The game had completely normal tessellation values in normal game view - but if you turn on wireframe mode, all culling and tessellation distance values turn off essentially. So the uninformed press that ran with the story thought a massive tessellated ocean was rendering at all times and barriers are tessellated to invisible amount of triangles... which is not what the game does at all in real time mode. Easily confirmable by using Renderdoc nowadays or just by turning on and off tessellation and measuring the performance on period correct hardware (which I could show on this forum if people are interested).

 

No, Crysis 2 has one of the most impressive implementation of Tessellation. Touching a huge amout of assets and modify them in a way which wasnt and isnt not possible today (or today with the UE5).

The same happens with RDNA2 today, too. Problems with Raytracing and ML will always result in a huge performance impact. So dont blame developers using modern techniques to render games and not relying on outdated ways like static geometry or SSRs...

 

Not agree, modern textures are up to 4K and any resolution limitations are visible only up close to the objects and if these object are large, while 99% of time during gameplay we don't stare at walls so that texels would be apparent.
There are also materials in modern games. Materials usually contain several normal, roughness and albedo maps - base textures for macro and detailed tiled textures for micro scale, so with materials you don't need bake in all micro details into 8K textures to get good detalizations up close to large objects.
Also, glossy materials, brushed and polished metals, paints and other materials have reflections on them and these reflections also add up to the detalization.
And the last, but not least, usually people don't stare at walls right in front of them, so texel density shortcomings are visible only in corner cases and avg scene's texel to pixel density with modern hi-res textures will always be close to 1 to 1 for the majority of screen area.

As for shadow maps, they suffer from the same mapping problem as all other textures, another reason to ged rid of them altogether since even VSM can't provide the 1 to 1 mapping in many cases (not speaking of tons of other limitations).

Such effects as shadows, AO, etc will always have better detalization when these effects done either in screen space or in world space via RT.

They are hybrid most likely and the sad part is that they are quarter res so they don't provide the 1 to 1 mapping either and the final image appears lower res as if it would have been with higher res options.

It seems you don't play games a lot because 1080p to 4K upscaling artifacts are quite noticeable for all effects be it particles, screen space or RT effects.
When blended in with high res lighting and textures, these low res effects create appearance of lower resolution due to how the spatial upsampling works (even depth aware).
There are many tricks to get rid of the low res look - checkerboarding with temporal accumulation in output resolution for example, but I have not seen these in FC6 or RE VIII.
Aside from the low res edges, reflections, shadows, and other artifacts due to upscaling, effects can't scale infinitely low in resolution since it would make glossy reflections, indirect shadows, AO, etc way less distinguishable and stable.

That's just remasters, not remakes. Crysis 1 remaster is built on an outdated codebase with all related single threaded limitations, that's not a devs' fault.

UE5 already does the same with lighting - it ditches lightmaps and swaps them to Lumen (this is one of the most praised things in UE5), which can use both HW and SW ray tracing. I can imagine in a few iterations they can simply switch to fully HW RT as HW RT is far more versitile (works not just for static).

Materials are usually simplified for RT. Metro for one doesn't even keep texture's UVs in BVH, 1 per object color is used for color bounce.

We are already on track. Most popular game engines already support BVH and RT, even the most graphically impressive first party console exclusives have impressive RT implementations and new games with RT appear every month.

 

Came to be known as Jersey Barrier Syndrome I believe

 

Indeed.

 

Me and countless others tried and correct this naive misonception to no avail, somehow the conspiracy theorists loved the idea of over tessellated concrete slaps and oceans, and the evil NVIDIA empire as If NVIDIA dictated what objects the developer should apply tessellation on.

 

Remember the infamous water supposedly not being culled in Far Cry as well? Same story.

 

Read Dictator's reply.
There never were anything in Crysis 2 which could've been considered "overtessellated".
GCN1/2 was just bad at tessellation, period.
It is in fact very similar to how RDNA2 is just bad at raytracing now.
And will most definitely end up the same way, with all the drama from the usual members here and elsewhere disappearing once AMD will get to speed with their RT h/w.

 

The truth is inconvenient so I doubt any minds will be changed this time either.

 

Btw - another easy confirmation of the fact that Crysis 2 did Not overtessellate anything is in the command r_tessellationmintrianglesize, which is 8 aka 8 pixels

 

I strongly doubt the drama has much or anything to do with directly AMD or their pc GPU's. While NV is ahead by alot now with turing and ampere on basically all levels, i think AMD might narrow that gap with their oncoming products. That and gpus from AMD like 6800 and faster arent that bad, direct comparisons to NV doesnt have to be made. As a pc platform gamer and user, theres just not all that much to create drama and wars around forums i think. (hell you can have the GPU of your choice but ok).

 

That's what filtering is for. But just because texels aren't apparent doesn't mean texture resolution is sufficient.

 

Demands to know the retailer PSman shops at

 

Yes those plus the count of rays generated by a hit/bounce with a material theoretically varies with sophistication.

 

Transparent reflections are a small part of everything a game renders. Improving them significantly does not necessarily guarantee a large improvement to the net visual experience. No one sane expects RT to fix everything, that's our point. There is an extremely vocal group here who get extremely upset when someone doesn't agree with their RT be all end all views. You also can't ever really separate the performance aspect until performance is no longer a concern.

 

8 pixel triangles sure looks like overtessellation for some cases. Especially in a game released 10 years ago.
Why would there be 8 pixel large triangles in concrete slabs?

 

LOL, and the cycle repeats again.

I just outlined how performance is not that much of a concern (step down from Ultra, use DLSS), but you keep ignoring these aspects and continue down the same circular argument logic.