Unreal Engine 5

    Quote from AlexM

    8 hours ago, AlexM said: Runs amazing on my mobile 2000 series nvidia card. I'm gonna test it on a 970 this weekend.

    You mean runs well for an “old” card or overall it runs well? I’ve heard of ppl struggling (15fps) with a serious rig.

    Are you running your own tests, like a small map instead the demo Epic included?

    Just asking cos I thought I’d download it and try, get maybe accustomed to the new UI - I would be only interested in practicing and learning how to make UI/HUDs if I was

    Quote from blackdog

    6 hours ago, blackdog said: You mean runs well for an “old” card or overall it runs well? I’ve heard of ppl struggling (15fps) with a serious rig.

    Are you running your own tests, like a small map instead the demo Epic included?

    Just asking cos I thought I’d download it and try, get maybe accustomed to the new UI - I would be only interested in practicing and learning how to make UI/HUDs if I was

    Yeah running my own tests specifically without nanite so I can get a feel for how Lumen performs. The example project isn't a good test because it has so many other features (nanite being the main one). I didn't write a test runner though, just used perception.

    Were the people you heard that had issues using the example project? Because if so that's probably not a good test for Lumen specific performance numbers.

    Quote from AlexM

    15 hours ago, AlexM said: Were the people you heard that had issues using the example project? Because if so that's probably not a good test for Lumen specific performance numbers.

    Think they’re just overall running the whole demo stuff. I mean I’d be surprised if it was running fine, always read that early builds run at 2fps or something, expecting that it would run like UE4 is naive.

    Quote from AlexM

    On 6/3/2021 at 2:49 AM, AlexM said: I'm gonna test it on a 970 this weekend.

    I understood Lumen supports 1070 and higher so it'd be interesting if you manage to run it at all right now

    Quote

    Quote Currently requires an NVIDIA GeForce GTX-1070 or greater card for performance in Early Access. Lumen has many options for scaling down but requires further development recommended for use.

    https://docs.unrealengine.com/5.0/en-US/Rend…platformsupport

    Quote from blackdog

    13 hours ago, blackdog said: Think they’re just overall running the whole demo stuff. I mean I’d be surprised if it was running fine, always read that early builds run at 2fps or something, expecting that it would run like UE4 is naive.

    Ok yeah that's not a good comparison because the demo stuff uses nanite. I'm specifically only talking about Lumen. As I said, I've already done some tests and the performance is not far off from UE4 with SS reflections and post GI enabled. So it's running well in the 60+ fps region for me.

    I think naive is a strong word. Consider the fact that the GI aspect of Lumen is more about taking advantage of a data structure that has only come into use in computer graphics in the last 5-10 years. That data structure (signed distance fields) is known for being able to do very performant spatial lookups. Compare that to the less efficient classical GI and the fact it's often supplemented with lightmaps in UE4 and you can see it's not so clear. The reason that Lumen GI is so exciting to me is not because it's just a new tech that's taking advantage of some new GPU chip. This is an implementation of a lighting algorithm that uses recently discovered, more efficient data structures.

    Consider the following

    - By not having lightmaps you have one less paramter variant going into your surface shader so you have less draw call churn due to less shader variants.

    - By not having lightmaps your memory usage is reduced. SDF's are more memory efficient than lightmaps

    - You no longer need classical light transmission data such as ambient cubes or spherical harmonics because you have no static lighting.

    - Reflection captures are also expensive on memory if you use that method for reflections.

    Path tracing is of course a concern because it can be expensive, it's not clear to me yet if that's only used for reflections or if it's also used for the GI. From my understanding reading the papers the GI component might only have to rely on SDF's. Regardless, screen space reflections are already expensive as it is and there's nothing stopping you from using Lumen for GI and reflection captures for reflections if you want.

    So ultimately, after more testing and time I'll have a better idea about perf but it's certainly not naive to expect this alternate algorithm will be performant and from my playtests already it appears that at least on a 2000 series card it is indeed performant and within the realm of UE4 performance with similar lighting/reflection features enabled. As for how it scales down, that remains to be seen.

    Quote from Mr.Yeah!

    8 hours ago, Mr.Yeah! said: I understood Lumen supports 1070 and higher so it'd be interesting if you manage to run it at all right now

    https://docs.unrealengine.com/5.0/en-US/Rend…platformsupport

    Yeah it will be interesting to see because from a technical standpoint it seems there's nothing preventing a 970 from running it. I'm guessing they just mean it won't have acceptable performance on a GPU that old. If I were to guess, the pain point is software path tracing. Assuming I can get it to run at all I'll compare Lumen GI + Lumen reflections vs Lumen GI + reflection captures and see what difference we get.

    I'll post back with my results once I give it a try!

    And here is another test where I disable lumen reflections. Sounds like the path tracing theory has some validity. Almost 33% increase in fps

    Technical talk on how nanite works. It's quite fascinating and it's fairly high level too

    External Content www.youtube.com
    Content embedded from external sources will not be displayed without your consent.
    Through the activation of external content, you agree that personal data may be transferred to third party platforms. We have provided more information on this in our privacy policy.

    Quote from AlexM

    On 6/4/2021 at 7:47 PM, AlexM said: it's certainly not naive to expect this alternate algorithm will be performant

    Mmh not sure if we got our wires crossed, what I mean with naive is that we’ve always (I think) seen that for any new generation of software we also need more powerful hardware.

    We’ve seen a slowing of this in the last decade(?), and in certain aspects we end up with underused hardware even, but –overall– you don’t expect to get something as complex as a game engine to run on par or better than the previous generation on the same hardware, or at least not unless you’re running the bleeding edge of available hw.

    Sure, I’m not surprised that some parts of the engine have massive gains thanks to innovation, i wasn’t contesting your tests by the slightest ;)

    by the way super interesting tests you’re doing, I got the gist of it but of course it mostly goes over my head

    Quote from blackdog

    2 hours ago, blackdog said: Mmh not sure if we got our wires crossed, what I mean with naive is that we’ve always (I think) seen that for any new generation of software we also need more powerful hardware.

    We’ve seen a slowing of this in the last decade(?), and in certain aspects we end up with underused hardware even, but –overall– you don’t expect to get something as complex as a game engine to run on par or better than the previous generation on the same hardware, or at least not unless you’re running the bleeding edge of available hw.

    Sure, I’m not surprised that some parts of the engine have massive gains thanks to innovation, i wasn’t contesting your tests by the slightest ;)

    by the way super interesting tests you’re doing, I got the gist of it but of course it mostly goes over my head

    I ah see. Gotcha. I still have hope that if you remove the old GI solution and replace it with Lumen that maybe you can hopefully get something close but you are probably right, it won't hit full parity performance-wise but it should hopefully look much better (otherwise what's the point :D ). I guess the question I want to answer is how much of a performance difference will there be.

    Anyways the truth is, I have a LOT more testing to do before I can be sure of anything, I'll keep posting back here with my results though. I have a modular scene I made in Unity that I want to finish so I can bring it into UE5 and see what happens with a more real-world scenario.

    On a side note I gave the Unity screen space GI solution (probably the closest thing they have to Lumen) a try on the weekend. It tanked my framerate from 80FPS to 20FPS :(Maybe I should do a Unity SSGI vs Lumen comparison next haha.

    So I've watched the whole Nanite presentation and am about halfway through the Q&A but I wanted to point out some interesting notes about it.

    GOOD : With Nanite you no longer pay a rendering cost for the overlapping part of two meshes.

    So what this means is imagine you have 2 very high poly rock meshes. Each one 1 million triangles. Now you place these rocks so they are 40% overlapping each other. In a traditional renderer you pay the entire 2 million triangle cost to render the meshes. If these meshes are Nanite-enabled then it's able to cull the overlapping triangles so you only pay for the non-overlapping, visible portions of the mesh!

    You no longer have to feel bad about having too much mesh overlap when doing kit bashing or cliff modeling!

    GOOD : Auto-LOD via triangle clusters

    If you take the above example you actually pay even less triangle cost than what you save from the overlap because Nanite essentially works in a way where they never want to draw more than 1 triangle per-pixel (since there's no point in having multiple triangles per-pixel, that's lost detail). So you actually save more than the just the overlap cost, the geo that gets rendered is roughly 1 triangle per-pixel MAX (if there's a flat surface then you can have the same triangle across multiple pixels).

    BAD : Does not work well with glancing or geometry that's close to other geo, but doesn't overlap

    So what this really means at the end is that Nanite is bad for stuff like grass and leaves where you have lots of small planes that are near each other but don't intersect. In fact in the UE5 example project everything is Nanite EXCEPT the foliage for this reason.

    Apparently if you make a big grass mesh, it has a chance of screwing up the LOD decimation that can cause your grass to look like a spider monster.

    Quote from Bunglo

    38 minutes ago, Bunglo said: how the fuck am i suppose to play with this on my gtx770 with restocks of modern GPUs nowhere in sight

    What’s the deal with pc parts right now? Just totally out of stock because of COVID/crypto and the situation isn’t resolving? I want to upgrade soonish

    @FMPONE getting a GPU at retail price is basically impossible. They sell out less than a second after they're in stock, which is basically never. People have bots buying them up and some people just get lucky with their timing. Not to mention all the tracking sites that are constantly refreshing inventories from newegg, amazon, etc. Cards that run $400-$500 retail, like an rtx 3060, are going for $1200+ thanks to scalpers taking advantage of the situation. I could probably sell my 770 and make a profit, which is fucking nuts. It's a combination of covid, crypto, and all the shit that's going on with China like embargos and what not being placed on them. There's simply not enough manufacturing power to keep up with demand. I think Taiwan is picking up some of the demand if I'm not mistaken, but Taiwan can only produce so many chips, still not nearly enough. CPUs are in a similar situation but it's not nearly as bad as GPUs. You might need to wait a few months to buy the processor you want if it's out of stock but getting a new graphics card? Maybe in the last quarter of 2022.

    Quote from FMPONE

    42 minutes ago, FMPONE said: What’s the deal with pc parts right now? Just totally out of stock because of COVID/crypto and the situation isn’t resolving? I want to upgrade soonish

    Higher general demand in all silicon sectors + crypto demand is not helping :E


    I was 'lucky' to pickup a 3090 for work with only $500 on top of the retail price just before the elon enduced crypto run. Right now it's about $1500 over the msrp to get one here.

Participate now!

Don’t have an account yet? Register yourself now and be a part of our community!

Register Yourself Login