Thank for the effort @Arkiras !
I was also complaining about the performance in 5.4, but couldn’t be sire about the scenario. I think performance was worse only in minimal scenes and/or with classic “settings” (no lumen nor nanite, nor vsms, etc), because I also notices improvements in heavu scenes with newest techniques. So I wasn’t able to confirm a veredict.
Anyway I’m doing a benchmark software, so I will try to post here my comparison too (5.3 and 5.5 when final released)
Great analytic approach. Much appreciated!
It’s weird though, how different the outcome from different projects seems to be.
One aspect that gives me some hope is the fact that your RHIT went down almost 50% (!!) between 5.4 and 5.5. In my/our case, the RHIT is the main culprit of the observed performance decay. So, it’ll be very interesting to upgrade to 5.5.
This example only shows extreme bad and bad performance. 30fps on a 3070 in 1920x1080 with one character is broken (5.1). 50fps is still bad for this (stylised like?) map, one character and a 3070. The only thing that’s surprising is that it is equal bad with 5.3 and 5.4/5.5. But overall bad does not help. Depending on the map I get decreased FPS from 100 (5.3) to 70 (5.5), 120 (5.3) to 70 (5.5). 110 (5.3) to 90 (5.5). 60 (5.3) to 50 (5.5). Some maps are a complete desaster and urge for VSMs (and in that case all meshes switched to Nanite) but this only improves it from very bad to still worse.
Overall I’m pretty sure that it’s some shadow issue as VMSs at least improves it a bit (but still does not fix it).
As I said, the performance is low specifically because the content for this level was never designed to be used with Nanite. It’s a worst case scenario, as Nanite’s performance is hindered by masked materials.
Assuming things should run fast because they’re stylized is also a bad way to think about performance cost, especially in UE5. Most of the costs are not dependent on the geometry of the scene and will scale similarly regardless of the style. Lumen isn’t any cheaper in Fortnite just because it’s “stylized”
None of this was really the point either, I just wanted to see how GPU performance had changed in a controlled scene across versions using a large scene.
My last set of comparisons, this time an interior, same settings as last time (Hardware raytraced lumen, hit lighting for reflections, everything is Nanite, etc., VT is enabled in project, but not used on assets)
Well over a hundred lights in this scene, I set them all shadowcasting to really stress the performance because I wanted to test Megalights. I used VSM for all local lights except in the Megalights capture, in that one they are using raytraced shadows.
Base Scene, hopefully this is realistic enough…:
Measurements:
Edit: Now that 5.5 is released, I tested it again with megalights, best result so far:
Final conclusions:
Only testing on my own hardware so this is a sample size of 1, and my focus was very much on GPU performance. There have been a lot of changes to physics and skeletal animation that could have big impacts on performance but I didn’t bother investigating that as I’m less familiar with that area of the engine.
Basically… your mileage may vary.
Anyway, thats it, time for me to say farewell to my 5.1/5.3 installs. As always, hope someone found this useful.
I feel stupid, but I have to admit that I started panicking when I saw this thread and when I updated my android project to 5.4.
It indeed became more sluggish. But it turned out to be only for non-shipping builds: Wait For Task taking too long in Development Build
You write that megalights are a big win. But your screenshots look like mega lights are fake because the fps in 5.4 without megalights are the same as in 5.5 with mega lights.
You don’t measure individual feature performance by using FPS… You look at the individual timings.
Megalights costs are substationally lower than the lighting costs of the traditional deferred lighting + shadow depth passes, even in 5.4. The fact that some of the performance gain is offset by other cost increases in 5.5p is not relevant to the conclusion that Megalights is indeed a big win.
Also worth pointing out that even if the performance were the same, ML achieves it while consuming less GPU memory as it doesn’t need to store shadow depths for every light.
Edit: Now that 5.5.0 is no longer in preview I added the megalights timings to my previous post.
Thanks for that hint. I’ve tested plenty of maps now from market-place examples. Starting with AbadonedFactory… and I stopped somewhere at Polar. Some examples had no difference. Some had incredible difference in Editor. But things changed a lot when run standalone. Like 5.3-editor 70fps, 5.5-editor 20fps, 5.3/5.5 standalone 90fps. So I’m not sure… I’ve stopped my 5.4 tests for now. This release was somewhat doomed already with the telemetry plugin from start anyway. I would take a further look into 5.5. Also while testing I’ve found out that market-place assets like DeepElderCaves differ a lot. With ShadowMaps everthing is fast in 5.3. With VMSs you run out of pool quick in 5.3 if you nanite just all those mushrooms and cave meshes and get worse performance than 5.5 with VSMs. If everything moves from normal ShadowMaps to VSMs then Megalights are a good addition as many small lights could cause issues with VSMs.
Have you had a chance to profile 5.5 just yet and what CPU/GPU are you setup with?
I’ve referenced your quote here.
Lumen GI and Reflections feedback thread - Development / Rendering - Epic Developer Community Forums
If you scroll up I pleaded for a flag to still be able to use DX11 SM5 SWRT cause DX12 was just a huge overhead simply by swapping the RHI … We got the flag coming in a 5.5 hotfix however the lumen engineer seems unconvinced of my stat output if you read further up in the thread.
The specific machine I benched was a Ryzen 5 5600, 32gb 3200htz, samsung M2 2tb, RTX 2080ti
But unlike the sage above I’ve gotten consistent better performance out of AWS G4s with Radeon V520, Steamdeck, GTX series paired with both intel and AMD - All running DX11 SM5 vs DX12 SM5
No, we’re still on 5.4. Our project needs some plugins that aren’t available for 5.5 yet.
I’m running a 3090 RTX with an Intel i7-12700K, 64 GB RAM
