I appreciate this video even more seeing how you're only using a 1080 so we get a realistic, sort of average performance idea of what many of us would experienced. It wouldn't quite be as useful if you were using a 3090 which most of us will not have. Thank you for this real world test!
however, considering both consoles use rdna 2, and UE has hardware accelerated ray tracing, the RT performance would be considerably better than demo'd here, as the 1080 doesn't have rt hardware acceleration
@Caltrop not compared to consoles - the ps5 equivalent graphics card is the rx 5700xt or 2070 super, both of which are more powerful than the 1080, and both of which have hardware ray tracing acceleration. therefore, as most games are developed with consoles primarily in mind, those are the specs they will target. Because they have RT acceleration, lumen will have far better performance than shown in this video, due to the 1080 used in the video not having hardware acceleration
@@dmrm2161 You realise every modern console has a gpu more powerful than a 1080, with hardware acceleration for ray tracing? So every game made for the current gen will be targeting that hardware in mind
I mean the fact that you were able to get good performance with 600B triangles, without a 3090, is more than enough room to make a great game. Using Unity, it struggles with 3M lol
@@JorgetePanete I feel like no matter what, Unreal engine will always be one step ahead of Unity, atleast for realistic scenes (most other things are a different story)
So in other words, this still-in-testing version of nanite has basically no problem using a trillion triangles in a single level. The only thing that really caused the system to fail a little bit was the power used to keep up with the instances and other memory issues. That is amazing 🤯 sure, this is probably a very nice computer, but just imagine what we could get to with another system on top of this that was able to support memory just as good as nanite supports geometry counts. This is incredible!
Currently nanites have some problems, mainly in generating dynamic scenes with plants. Thare is a reason why every nanite demo is on desert or in town :P
@@PanSkrzynka_ Ofc, they said it only works with static meshes. It is a dynamic LOD system after all. Something that is animated and or deforms wouldn't be able to use the dynamic system properly. But if nanite takes a lot of the stress out, one is able to give more geo to plants.
Decided to run the stress test myself. With a 1050 ti and lumen off, i was sitting at a constant 30-35 fps, even while moving, and got up to 50k meshes before i decided to stop. with lumen on it can vary between 10-20 fps, so its not the prettiest, but its about the same if not better than trying to run Minecraft shaders. This test and the past two videos were a great insight into how nanite works and what makes it so great, really happy i was able to do it myself, so thank you!
Thank you for sharing your results! It's helpful to see how it performs on a 1050ti. Lumen is especially demanding, so it will be interesting to see what games using it do to keep frame rates up for older GPUs.
Good God. As a lifelong artist, and 3d artist for 25 years, including 10 years in games, Nanite is a paradigm shift. So.e enterprising person needs to make a sophisticated, quality, procedural spherical, planetary terrain generator that makes unique, playable worlds.
8:46 there is siggraph presentation about nanite on youtube, this is probably because their culling uses some data like z-buffer from last frame for partial occlusion culling, if that doesn't remove much triangles, their second occlusion pass is not worst case optimized thus the huge overdraw..
The amount of geometric detail possible is absolutely insane and will change game development not just for realistic games but more stylized ones too. Lumen also impresses me with its performance on GTX cards despite having no ray tracing acceleration hardware on board.
@@crestofhonor2349 Most of you guy's are not aware of what is truly going on right now but you will understand next winter. Psychopaths are running this planet unfortunately for us the peasants!
I can not even begin to comprehend how they can do such a complex seeming algorithm in *realtime*... without any significant pre processing. I suppose they must do a *bit* of preprocessing when you hit "apply changes", but still, it's unbelievable.
There is definitely pre-processing happening to build the Nanite data structures on mesh import (or when you enable Nanite). These geometry trees are then used at runtime to quickly swap out clusters.
@@LivelyGeekGames Oops, I didn't see that reply! That makes sense. I suppose the reason I assumed there was no significant preprocessing is because in that UE5 tech demo from 2020ish, the spokespeople were saying something along the likes of "you can import your models straight from ZBrush with *no baking*".
Somehow I have a feeling that NVIDIA and AMD will internally freak out about UE4 as it seems it will make newer and faster hardware basically obsolete for a few years. I mean. I use a GTX 1080ti that can keep up with most modern games at 1080p without any problem and usually my 7700K is the bottleneck. But with UE5 it feels like there is no scaling limit anymore on any game environment and it is mostly only depending on screen resolution but not scene complexity. Of course there are still other things that won't scale that well like physics and light but that's something thats both things that are already at a nearly unrecognizable difference level.
AMD will be happy because it looks like the new limit is RAM as the new hardware bottleneck But it also means that games can ship with movie level assets which will be 100s of GB instead of the smaller compressed ones. But also games only need to ship 1 LOD so it does cut each shipped object by like half or less
I bet lumen would perform better in a medium range RTX card but it's quite incredible the power that the engine gives you with geometry, is this for real the end of the LOD age?
Well nanite only works on static meshes, characters, plants and other meshes that deform, need LODs still. But yeah, it might be the end of LODs on static meshes.
I wish there was a comparison video. comparing regular LOD's to nanite. Nanite's performance is incredible, but compared to regular LOD's does it still have the same result? Also doesn't nanite require a higher base triangle count to render the same model? Where a artist might make some triangles very large due to not needed a lot of detail in that surface area, nanite might add several due to an edge it needs to cover. This is just speculation though. But I for sure would love to see a comparison video in which number of actors (instanced) is the benchmark score.
Running this test on a laptop computer with a 2060 with ActorsPerMesh >=1 : 25 fps at 10 000 meshes, it drastically falls each time you press b (from 0 to 10k) and the material doesn't load with ActorsPerMesh =0 : 25 fps at 90 000, meshes material properly set Also, animated mesh seems to work (static mesh + animation texture), all of them playing the same animation at the same time, i am around 40k entities (each mesh is around 2800 polygons) but it's hard to keep the triangle correct (deformed models). Might not be fully suitable for this Nice test project, I really enjoy finding C++ project like this, thanks
I have a question: When i add a Terrain Mesh from Blender to Unreal Engine, i need to setup a collision for it, but the default terrain in UE5 has no collision. So why extern meshes need a collision in UE5?
Alot depends on your gpu. I tried millions of polys with nanite but all it done was eat my gpu and all ypu get is exhausted vram issue and i have a 8g rtx card
I find this video hard to believe because I have a way more expensive graphics card and can’t even get 120 fps using a brand new empty project. My guess is you disabled a lot of things to get nanite to run quick.. You mentioned lumen for example, but what is the point if we are never going to use nanite without lumen?
Do you really only using GTX 1080??? I'm using 2070 super and just using 2 object low poly, it's already only 80 fps......... Anyway, really really cool experiment!
This basically means that triangle count is now irrelevant, and it's more of the technical details and rendering makeup that actually affects performance.
Couldn't you animate a nanite mesh if all the body parts were being scanned separately by unique cameras from a source of placed assets but the camera displayed the view as "moving joints on a projection?" Like if you placed a foot piece flipped the camera as if it was simulating it walking but it was the camera moving the whole time would that equal a way to cheat a nanite object to be perceived as animated? Or go simple create a static mesh wall and have a camera move up and down from their then copy that movement as a still camera and possibly create an animated door.
You can certainly move the static meshes around or even create rigs and assign static meshes to the bones and animate them that way. You just can't deform the meshes yet (like skeletal meshes or world position offset in materials). Hopefully they add support for this in a future version of Nanite.
@@LivelyGeekGames I hope so too. But still. I was thinking like how george lucas animated ships could you create a simple moving object simulation in a video like that? Using multiple cameras then recreate it. Because if. You can rig enough moving nanite functions aka moving object like a robot. Or a door. It would be cool to show that to people. Or give developers new ideas to implement nanite into their games.
The PS5 GPU is around the same power in rasterization compared to the 1080ti..I think once devs get to grips fully with ue5 we are gonna see some great things as this is doable with even console level hardware.
Thanks a lot - I must also say thanks that you FIRST OF ALLL tell what sort of hardware you have - That tell me that with a good computer - not only with a ridicouslsly fantastic computer - you can really get fantastic performance ! Thanks a lot ! Kind Regards !
Unless you're making a game with 1 million snowballs, its not a very practical test. Seeing how a few to several thousand unique meshes perform would be a better test.
the problem there would be the huge amount of memory needed to store the geometry, it's not much of a nanite limitation but a "rest of the engine and hardware" limitation
Yeah, for sure. It is nice having the high detail options when you want it though. Makes it much easier for 3D artists to just create and not worry about poly counts.
really wish i knew what ur trying to convey with actors ... u say 10 actors per mesh at one point.... seems backwards. Arent mesh actors inside a blueprint? Arent meshes actual actors... do u have your terminology wrong??? Are you trying to say 10 actors instanced within a blue print? so confusing how you word things.
if there was 10 types of meshes, it would be 10 times less optimized, 1000 types of same object, not same as 1000 types of 10 objects, this test not objective and does not show how nanites would work in the game (optimization will be like in Remnant 2)
I appreciate this video even more seeing how you're only using a 1080 so we get a realistic, sort of average performance idea of what many of us would experienced. It wouldn't quite be as useful if you were using a 3090 which most of us will not have. Thank you for this real world test!
however, considering both consoles use rdna 2, and UE has hardware accelerated ray tracing, the RT performance would be considerably better than demo'd here, as the 1080 doesn't have rt hardware acceleration
Makes me want to try this on my computer so I can test it with a 3090
@Caltrop not compared to consoles - the ps5 equivalent graphics card is the rx 5700xt or 2070 super, both of which are more powerful than the 1080, and both of which have hardware ray tracing acceleration. therefore, as most games are developed with consoles primarily in mind, those are the specs they will target. Because they have RT acceleration, lumen will have far better performance than shown in this video, due to the 1080 used in the video not having hardware acceleration
"you're only using a 1080 ". You are a joke.
@@dmrm2161 You realise every modern console has a gpu more powerful than a 1080, with hardware acceleration for ray tracing? So every game made for the current gen will be targeting that hardware in mind
I'd really like to see this done with more complex textures to see how that effects this instead of a simple white texture
I mean the fact that you were able to get good performance with 600B triangles, without a 3090, is more than enough room to make a great game. Using Unity, it struggles with 3M lol
Some guy is making its own Nanite on Unity 👀
@@JorgetePanete I feel like no matter what, Unreal engine will always be one step ahead of Unity, atleast for realistic scenes (most other things are a different story)
@@DylozWitty tbh you could use Unreal Engine and try to make a 2D game and be able to make a billion triangles with sick graphics and a good 2D game
So in other words, this still-in-testing version of nanite has basically no problem using a trillion triangles in a single level. The only thing that really caused the system to fail a little bit was the power used to keep up with the instances and other memory issues. That is amazing 🤯 sure, this is probably a very nice computer, but just imagine what we could get to with another system on top of this that was able to support memory just as good as nanite supports geometry counts. This is incredible!
Currently nanites have some problems, mainly in generating dynamic scenes with plants. Thare is a reason why every nanite demo is on desert or in town :P
Instances are the new triangles
@@PanSkrzynka_ Ofc, they said it only works with static meshes. It is a dynamic LOD system after all. Something that is animated and or deforms wouldn't be able to use the dynamic system properly. But if nanite takes a lot of the stress out, one is able to give more geo to plants.
the new UE5's ECS (Mass) is like Unity's DOTS, and was the missing part of the puzzle
@@farrex0 their plan is "Nanite everything", I hope to see that some year
Wow, I’m mind blown. Nanite is OP.
The real stress test video I was looking for... Good job!
And you pushed the lighting quality! Thank you.
Decided to run the stress test myself. With a 1050 ti and lumen off, i was sitting at a constant 30-35 fps, even while moving, and got up to 50k meshes before i decided to stop. with lumen on it can vary between 10-20 fps, so its not the prettiest, but its about the same if not better than trying to run Minecraft shaders. This test and the past two videos were a great insight into how nanite works and what makes it so great, really happy i was able to do it myself, so thank you!
Thank you for sharing your results! It's helpful to see how it performs on a 1050ti. Lumen is especially demanding, so it will be interesting to see what games using it do to keep frame rates up for older GPUs.
That was an entertaining breakdown. Cheers.
Thanks!
Thanks for the in depth testing! Especially with all the combinations of duplicated mesh types. Very useful information.
Good God. As a lifelong artist, and 3d artist for 25 years, including 10 years in games, Nanite is a paradigm shift. So.e enterprising person needs to make a sophisticated, quality, procedural spherical, planetary terrain generator that makes unique, playable worlds.
Nanite will be a bigger leap than DLSS or RTX.
Is that what they call magic?
This is incredible. Math/Code trickery of this high level should have high amount of glitches in that scene, but no.
Incredible engine, and incredible content ! keep it up !
Thanks, will do!
Holy crap! This is amazing!
8:46 there is siggraph presentation about nanite on youtube, this is probably because their culling uses some data like z-buffer from last frame for partial occlusion culling, if that doesn't remove much triangles, their second occlusion pass is not worst case optimized thus the huge overdraw..
That's very impressive. Asset modelers would focus more on textures now.
The amount of geometric detail possible is absolutely insane and will change game development not just for realistic games but more stylized ones too. Lumen also impresses me with its performance on GTX cards despite having no ray tracing acceleration hardware on board.
Too bad they are going to destroy the world!
@@petertremblay3725 ??
@@crestofhonor2349 Most of you guy's are not aware of what is truly going on right now but you will understand next winter. Psychopaths are running this planet unfortunately for us the peasants!
@@petertremblay3725 what in the world does that have to do with my comment on UE5
@@crestofhonor2349 Well wise guy how will you enjoy technology if society is getting destroyed?
I can not even begin to comprehend how they can do such a complex seeming algorithm in *realtime*... without any significant pre processing. I suppose they must do a *bit* of preprocessing when you hit "apply changes", but still, it's unbelievable.
There is definitely pre-processing happening to build the Nanite data structures on mesh import (or when you enable Nanite). These geometry trees are then used at runtime to quickly swap out clusters.
@@LivelyGeekGames Oops, I didn't see that reply! That makes sense. I suppose the reason I assumed there was no significant preprocessing is because in that UE5 tech demo from 2020ish, the spokespeople were saying something along the likes of "you can import your models straight from ZBrush with *no baking*".
formative video experience mr. lively geek!!! xo
Somehow I have a feeling that NVIDIA and AMD will internally freak out about UE4 as it seems it will make newer and faster hardware basically obsolete for a few years.
I mean. I use a GTX 1080ti that can keep up with most modern games at 1080p without any problem and usually my 7700K is the bottleneck. But with UE5 it feels like there is no scaling limit anymore on any game environment and it is mostly only depending on screen resolution but not scene complexity. Of course there are still other things that won't scale that well like physics and light but that's something thats both things that are already at a nearly unrecognizable difference level.
"this one element of one particular game engine works really well so now new hardware is obsolete"
@@bearwynn, yeah, textures will be a new problem now, but he's kinda right....
AMD will be happy because it looks like the new limit is RAM as the new hardware bottleneck
But it also means that games can ship with movie level assets which will be 100s of GB instead of the smaller compressed ones.
But also games only need to ship 1 LOD so it does cut each shipped object by like half or less
I bet lumen would perform better in a medium range RTX card but it's quite incredible the power that the engine gives you with geometry, is this for real the end of the LOD age?
Well nanite only works on static meshes, characters, plants and other meshes that deform, need LODs still. But yeah, it might be the end of LODs on static meshes.
It's soo great ! Thank you to do this content :)
I'm glad you found it helpful!
102 fps on a 3090 with nanite and this test.
Nice, and that was with 8M meshes?
Lumen on or off ?
Because it works with caching on so many different levels...going fast might have spiked the triangle count. A lot of stuf was "new" probably?
Yeah, that's what I was thinking. Cold cache and it had to repopulate, and for some reason it needs a lot more triangles than there are pixels.
I wish there was a comparison video. comparing regular LOD's to nanite. Nanite's performance is incredible, but compared to regular LOD's does it still have the same result? Also doesn't nanite require a higher base triangle count to render the same model? Where a artist might make some triangles very large due to not needed a lot of detail in that surface area, nanite might add several due to an edge it needs to cover. This is just speculation though. But I for sure would love to see a comparison video in which number of actors (instanced) is the benchmark score.
I feel scared of them, *playing Minecraft with 20 fps, and then there start to render few millions of entities*, i afraid of them
Running this test on a laptop computer with a 2060
with ActorsPerMesh >=1 : 25 fps at 10 000 meshes, it drastically falls each time you press b (from 0 to 10k) and the material doesn't load
with ActorsPerMesh =0 : 25 fps at 90 000, meshes material properly set
Also, animated mesh seems to work (static mesh + animation texture), all of them playing the same animation at the same time, i am around 40k entities (each mesh is around 2800 polygons) but it's hard to keep the triangle correct (deformed models). Might not be fully suitable for this
Nice test project, I really enjoy finding C++ project like this, thanks
Can you do this again but with the new update with foliage?
you are so underrated
amazing
like Limit Breaker for Unreal
What are the Vram usages like with nantite vs LODs?
I'm not sure, I didn't setup LODs and measure that. You could download the test project I made and run some tests pretty easily though.
I have a question: When i add a Terrain Mesh from Blender to Unreal Engine, i need to setup a collision for it, but the default terrain in UE5 has no collision. So why extern meshes need a collision in UE5?
Excellent video!
absolutely..... unreal ;)
love the vid what are your computer's specs
Even 1 quadrillion triangles... :)
Well, attempting to work with 1 trillion triangles helps me puts the 29.8 trillion US National debt into perspective… 😳
😳
Alot depends on your gpu. I tried millions of polys with nanite but all it done was eat my gpu and all ypu get is exhausted vram issue and i have a 8g rtx card
If there is a way to save memory then it would be great.
Great test!!
I find this video hard to believe because I have a way more expensive graphics card and can’t even get 120 fps using a brand new empty project. My guess is you disabled a lot of things to get nanite to run quick.. You mentioned lumen for example, but what is the point if we are never going to use nanite without lumen?
Does this run any better with version 1.0? With Nanite and also with Nanite and Lumen.
Not sure, I've not tried it with the new 5.0 release yet!
Do you really only using GTX 1080??? I'm using 2070 super and just using 2 object low poly, it's already only 80 fps.........
Anyway, really really cool experiment!
so can you use this with pcg as well to get better frames
This basically means that triangle count is now irrelevant, and it's more of the technical details and rendering makeup that actually affects performance.
Couldn't you animate a nanite mesh if all the body parts were being scanned separately by unique cameras from a source of placed assets but the camera displayed the view as "moving joints on a projection?" Like if you placed a foot piece flipped the camera as if it was simulating it walking but it was the camera moving the whole time would that equal a way to cheat a nanite object to be perceived as animated? Or go simple create a static mesh wall and have a camera move up and down from their then copy that movement as a still camera and possibly create an animated door.
You can certainly move the static meshes around or even create rigs and assign static meshes to the bones and animate them that way. You just can't deform the meshes yet (like skeletal meshes or world position offset in materials). Hopefully they add support for this in a future version of Nanite.
@@LivelyGeekGames I hope so too. But still. I was thinking like how george lucas animated ships could you create a simple moving object simulation in a video like that? Using multiple cameras then recreate it. Because if. You can rig enough moving nanite functions aka moving object like a robot. Or a door. It would be cool to show that to people. Or give developers new ideas to implement nanite into their games.
@@Keepit10011 Yeah, it's very easy to move static meshes around with Blueprints or C++.
Try it in Unity and see what happens
Is it not the fact, that there are no trillion polygons? They are nit displayed altogether, only in the area where spectator is.
Yeah, Nanite culls out all the unnecessary triangles to only show what's really needed.
The PS5 GPU is around the same power in rasterization compared to the 1080ti..I think once devs get to grips fully with ue5 we are gonna see some great things as this is doable with even console level hardware.
Thanks a lot - I must also say thanks that you FIRST OF ALLL tell what sort of hardware you have - That tell me that with a good computer - not only with a ridicouslsly fantastic computer - you can really get fantastic performance ! Thanks a lot ! Kind Regards !
amazing
Unless you're making a game with 1 million snowballs, its not a very practical test. Seeing how a few to several thousand unique meshes perform would be a better test.
You needed to download some rams.
Try instancing the meshes
ruclips.net/video/oMIbV2rQO4k/видео.html
I do switch between using non-instanced and instanced static meshes as part of the performance comparison.
Very impressive but i think this would not be possible with all unique meshes.
the problem there would be the huge amount of memory needed to store the geometry, it's not much of a nanite limitation but a "rest of the engine and hardware" limitation
Matrix version. 0.05.
I would find more basic games more interesting to play on. You dont always need HD
Yeah, for sure. It is nice having the high detail options when you want it though. Makes it much easier for 3D artists to just create and not worry about poly counts.
really wish i knew what ur trying to convey with actors ... u say 10 actors per mesh at one point.... seems backwards. Arent mesh actors inside a blueprint? Arent meshes actual actors... do u have your terminology wrong??? Are you trying to say 10 actors instanced within a blue print? so confusing how you word things.
Hi
if there was 10 types of meshes, it would be 10 times less optimized, 1000 types of same object, not same as 1000 types of 10 objects, this test not objective and does not show how nanites would work in the game (optimization will be like in Remnant 2)
hehe
boring test, same object a zillion times is a case for instancing
Watch the whole video. He addresses that.