Nanite Performance is Not Better than Overdraw Focused LODs [TEST RESULTS]. Epic's Documentation is Dangering Optimization.

It looks like the voxel part still needs some clean up work (lighting, and the shadow artifacts in places) - I wouldn’t be surprised if it gets faster though, so will hopefully be more comparable to the 6% triangles. Hopefully the amount of resources they use comes down too.

1 Like

Yeah for sure the main problem is that with distan color and shadows, cause its nice reduce geometry and overdraw.

EDIT About memory
propably 2 instances of the game were in buffer (that crashed didnt close)


Also command

r.Nanite.MaxNodes 5000000 - fixes glitches :innocent:.

1 Like

I too seem to be having good results with nanite and vast amounts of grass. I couldn’t make vast-expanses with billboards and have it still cast accurate shadows at a distance, or just the sheer-number and reach of being able to have a mesh vs a square.

Pics

I have an old machine, 10y/o proc + 4060 RTX w8gig, 60fps is good results for me at 2k res with dynamic everything + fog and atmospheric effects (UDS). This is also 554, 56 borks the landscape so I cannot reliably test..

There’s another Quinn mesh in that 2nd pic, but good luck finding it, it’s faaaar away and still not as far as the grass goes..

Video

https://www.youtube.com/watch?v=QI_1jyqBE2k

2 Likes

Looks Great and Nice Frames i really like atmosphere though - from Conan The Barbarian(1982) like a scenery from Tabernas!

Also wanted report that yesterday but couldnt through submission ticket.. wait like 30 mins and not proceed. Iam turning now HWRT off and will see if it crash again.

Update: Hardware Raytracing is Unstable IMHO.

Software Stable and how Streaming get hickups with too much going on (Character speed :D)
iam sad cause Hardware RT from 5.5 is real thing.

Please bump if you want better engine :slight_smile:

1 Like

Misson Accomplished. Open World and Nanite Discovered.
Its not possible atm with Unreal Engine as 5.6 (without nanite assemblies) to deliver a Game that have milions of poly foliage, its seamless (8k level size+) not consume 10+ GB of memory runs 60fps , have that infinite view, load below 1 minute and not hickup with World Partition.
Also you cant do that too with LODS^

I get to the point with that science and effort
iam able to plant foliage on all of my terrain , have good loading time, nice close looking foliage.

found that sweet spot - balance between streaming/visuals to not have hickups and that works without new streaming system! (experimental) .

Look solutions to reduce as much triangles and memory for assets mostly to have more dense foliage.


For other Unreal fans:
i made video of course with propably best thing i could do for current gen with infinite visibility for large scale enviroment.

If anyone disagree with that- deliver then guide
stop making benchmarks in editors without checking resources , possibility to package and running .exe - no matter from each coast You are :heart_hands:
All The Best
myasga

2 Likes

There’s also the use of occlusion which will help a great deal here on render times, most of the time you wouldn’t see so far away in the distance, having things like rocks, gullies, buildings etc around - hiding whats behind it.

2 Likes

I agree but that Would be fun?;D I can reduce loaded hlod like twice of the map ,apply height fog to reduce memory that is loaded to the world ,include assets culling. There are ways . I did all with base tools now going to check some external. Questions is PCG allow to plant more?, New streaming system? Meshpack or a Level streaming optimizer? Can i handle put assets in same position that are baked on HLOD With Yours RDInst? Or IT will allow me to include on HLOD?.That’s new goals.

2 Likes

I think all those systems allow baking to HLOD - streamed PCG tiles are fast but I don’t know how they would bake to HLOD - it should be possible.

Yeah I was going to mention height fogs too, but I think the occlusion is where Nanite shines, it’s where you can overtake traditional rendering to get much more on screen - as long as your meshes have fine polygons to allow good cluster culling, that’s the important part.

1 Like

Just to comment on level-of-detail… I can see the crack in this little-tiny-rock on the left, and it still has an actual crevice in the thing, mechanically, and insofar as lighting… Crazy that it’s texture-driven and comes out that well.

Yes, displacement eats the FPS but another reason to want to try in 561…any-day now, Epic..

Why is that crazy? ANY DCC has been doing it correctly and procedurally for 15 years at least. Unreal is the one well behind the curve - generating a mess by adding details with tessellation, ignoring best practices, etc.

Behind? Likely, but insofar as what I can produce, accessibility of the product, and honestly: quality of the output.

Any dedicated app is going to surely be better at the-thing vs unreal, agreed. As well, any static/baked mesh surely will be more performant but I am unsure how I could even begin to use a baked-mesh at this level of fidelity. I’d likely need displacement on top of it, so why not just all on the landscape.

And I don’t even need a landscape so much as a contiguous-surface, be it wrapping up walls and then flattening out to be the floor; I’d do it all in the shader. There’s no seams b/c it’s all the same shader. :smiley: It works for my purposes and tested on a newer machines that easily get 90’s for FPS (my proc is a major bottleneck with only 512MB L1 cache…). So performance is not a major concern.

Hard to tell how bumpy it is when not in motion

You got the idea, and that’s kind of the point - instead of fixing the landscape system (which they are incapable of because they lack imagination?) They are attempting to use Nanite as a “cure all”.
Its a good theory - but its the epic team. They succeed at selling fortnite skins to kids. Not at making the engine work… and sad as it may be its a proven fact by over 6 years of constant failure…

And to be honest I’m truly surprised you are able to get half decent results. Good work :wink:

Hey,
cross fingers that its only 5.7 issues and not killed my hardware ;(
FastGEO - i could only record once.exe - file crash often.
Nice improvement vs my older 5mln it really helps to reduce that huge packets impact = sad that its not replcated..


Will try to get some more on 5.6 and if it will crash with previous scenarios [*] my PC or component.

All the best!

1 Like

Driver updates can be causing this too - Nvidia is often guilty of screwing us over (similar but not worse than Epic).

If you know that’s it, and you happen to know what driver version was working before, with a bit of work you could revert to it…

3 Likes

thank You for hint,NV dropped just newest yesterday - maybe it can help with stability.
EDIT: CantPackage with newest drop of engine oopsie.

I was able to at least test my landscape-solution in 561 (rebuilding the nanite landscape still makes pieces go missing) but the upgrade into 561 worked and I was able to run the standalone.

~25% increase in performance 554 vs 561. Even better, the dips in fps were less impactful and increasing the density of landscape-grass and/or the cull-distance cost much-less per unit increase vs 554. Bugs aside, a decently-solid increase in performance.

554

Looking out, on that mesa/plateau, the bright spot top/center is another Quinn mesh; 45ish fps

~55fps looking down

561

~55fps

Straight-down, ~72fps!

Increasing the cull-distance for grass from 1 to 2 (double the radius) only cost me ~1fps in 561 vs several/10+ in 554, so another solid-gain.

It might still not be optimized, but a good-gain is a good-gain.

NOTE: the rendering didn’t change, between versions, I just futzed with the color between tests.

1 Like

Hey!, someone decode Nanite Assemblies going to do test asap

Liam Wedge has done an in-depth analysis of them on artstation.

4 Likes

I made a performance-pass on my setup and I was able to get ~5ish FPS increase across the board, plus make my grass look better, and better color-grading across the board.

561 - new results

~70fps:

grass out to ridiculous-levels (1.5km):

NO pop-in, no LOD-ing, no artifacts of any sort in that arena; there’s a certain seamless-axis across distance where nothing jumps-out in that regard… With my materials, with respect to what Lumen likes/dislikes insofar as extreme-roughness, etc; there is, and I mean almost none, practically none of that dark-spot popping with foliage, etc. The image is stable even without DLSS. MUCH BETTER in 561 vs 554, I’ll give Epic that, along with the performance-boost 56x is a winner, despite the bugs. You know those will be worked out, but the updates to the core-rendering part of the engine paid-off for my experience.

Also, I wasn’t trying to be kind to the setup.engine, I bumped a few settings in UDS, etc:

Summary

Material is NOT cheap, but I’ve been able to reduce the cost (texture-samples, LWC conversions/usages) as much as I can. Increases/reductions in the instruction-count don’t really add/subtract many frames; I’m not material/GPU bound here:

Summary

EDIT: more to the point of the thread, I am not expecting the engine to save me from myself and the (unreasonable) demands I might make of it, but I CAN say, what I am able to get away with now, at a decently-performant/playable level even, LODs wouldn’t let me do it as well or as seamlessly. No, it’s not going to cost the-same as LODs but it also does different things for me, like ensuring a smooth, regular, unpoppy deformation, every-time.

EDIT2: and as far as detail, it’s crazy how small it can get, being it’s shader-driven, self-occludes (the lighting is great), and is performant-enough to be 60ish fps on a modern-machine (mine is a 10 year old CPU…).

Landscape Displacement

An ostensibly-flat surface deformed in the shader, notice how small the shadows and details get… And for the jaded b/c I know there are haters, this is all using temporal-blending; no blurring in motion. That’s also across 4 different texture.material sets (rock, rocky, small-gravel/sand, grass.

1 Like

Now this guy is telling people to review bomb UE5 games because “UE5 = Bad”…

@TheKJ You have not ONCE showcased REAL knowledge of anything relating to game development nor have you tried building something that is worth anything to showcase this boogie man version of UE5 that you’ve created in your head. You crap on people like me who have built real games with UE5 and have actually found ways to work with these new technologies to get them to perform extremely well even on low-end hardware.

What you’re doing is spreading B.S lies and encouraging people to punish developers who are actually putting in the work, just because you personally don’t understand how the technology functions. That doesn’t send a message to Epic… it only hurts studios and teams that are pouring years of effort into their games.

If you actually cared about performance, you’d spend your time learning the engine, experimenting, and contributing solutions instead of calling for review bombs. Until then, your arguments have zero credibility, because unlike you, many of us are already proving that UE5 can perform beautifully… even on 2019-2020 consoles at native 4K resolution.

These are the settings I used on the XSX and YES, TSR, VSM’s, Lumen and Nanite are all used at NATIVE 4K.

5 Likes