UNREAL ENGINE 5.4 EXTREMELY BAD PERFORMANCE

Your RHIT climbed from 5ms to 25ms. I encountered the same issue with UE 5.4.4.

For some reason the RHIT is no longer parallel to the CPU and GPU workload, but only picks up work once CPU and GPU are done. No idea what’s going on, but it cut my performance in half.

I’ve posted my UE Insight profile here:

I’ve older UE profiles from previous UE versions (i.e. pre 5.4.4) where this problem did not occur. The RHIT was parallel to GPU and CPU. Definitely a massive engine issue.

Funny enough, renderer parallelization was announced a big thing for 5.4. That’s when it started to become a problem on my PC.

5 Likes

For anyone who’s unsure how to read an UEInsight profile, I’ve marked the issue in the shot below.

As you can see, most of the time (i.e. more than 50%), CPU and GPU are just idling, waiting for the RHIT to finish.

You can have a CPU and GPU as fast as you want, if UE doesn’t let it work, it’ll literally render your advanced hardware useless.

The problem does, however, not occur on all machines. In our dev team, only machines with a 3090 RTX are affected. AMD GPUs seem to be immune to the effect. However, this might be down to particular AMD models etc.

But the overall issue is so heavy that I consider the current build of UE as not production ready, even for basic projects. :frowning:

Hope someone from Epic sees this and realizes the critical severity of this issue.

11 Likes

MAN RUN A PROJECT YOURSELF ONE IN 5.2 and one in 5.4 and you’ll see the performance issue. WHAT THE HECK ARE YOU DOING AT EPIC. How are we supposed to build games with your busted engine, fix your goddamnnn engine it’s becoming increasingly unusable!!!

3 Likes

haha, I laughed at the comments))
p.s. but it’s all really sad of course(

update: when the first version 5.4.0 was released i compare the fps on my project at the same settings between 5.4.0 and 5.3.2, fps in 5.4.0 version was really bad, than in 5.3.2
But for now, i downloaded 5.4.4, disable Telemetry plugin and compare again between 5.4.4 and 5.3.2, it’s strange, but fps are the same! I have 2070 super rtx + intel i7-8700k overclocked :slight_smile:

It seems like not every GPU is affected equally. A critical indicator seems to be the RHIT. If RHIT is low, chances are, your GPU is good with 5.4.

2 Likes

So. had anyone tested 5.5 in this regard? Do we have some hope or are we all doomed?

5.5 preview is extremely bad performance as well of course. You can choose to stay with 5.3 forever or search for an alternative. It seems like nobody really cares.

1 Like

It’s no GPU issue (tested with 4070 and 4060Ti). The difference is only the engine version.
5.4/5.5_pre r.nanite 1 → 40fps
5.4/5.5_pre r.nanite 0 → 90fps
5.3.2 r.nanite 1 → 110fps
5.3.2 r.nanite 0 → 110fps

Funny thing is there is no nanite mesh anywhere so I’ve no clue why this setting even affects something in 5.4+. It’s a massive downgrade in FPS (even if turned off its a lot worse) and almost a half year now and plenty of hotfixes that changed nothing. It was also really predictable now that nobody cared about all those begging for performance but we get a bunch of new features (that you can’t use if you can’t upgrade).

The telemetry thing is just one of those issues that are easy to get ridd off but there are a lot more major issues with 5.4+

1 Like

Anyone else have what seems like a memory leak in UE5.4?

Been constantly getting crashes, both the UE engine and from explorer to blue screen. Left task manager open to see wth is going on and to my surprise, even on a blank project, my ram topped at around ~29GB/32GB and the engine crashed. This happens every 20 or so minutes. Starts to stutter then crash. I’ve upped my virtual memory to 128gb, after the rams full the virtual mem gives me another 10 or so min before crashing but then the whole pc is retarded slow.

Specs:
CPU: Ryzen 5800X3D -Water cooled
GPU: 3080TI -Water cooled
Ram: 32GB same kit ram (corsair 3200 cl16)
Motherboard: Tomahawk Max 2
SSD: 2TB NVMe

As you can see, my system is quite capable. I’ve messed around with UE on worse systems before and even before I upgraded my 16Gb ram to 32Gb, I’ve never had this problem. I’m literally forced to restart the engine every 20 or so min. Every time I click on the play icon, the ram usage goes up by a few 100mb but never go down when I hit Esc…

I have not, no. I recently upgraded to 5.4.4 and have been working pretty-crash-free, actually.

I’ve even left my machine on overnight w/open project and it was still there. 4060 here if that matters.

Hell idk, I’ve even used the epic launcher to verify the engine. Modified the garbage collection within the .ini file but still… idk, idk

Glad that you at least don’t have the problem lol it’s a headache…

I’m planning on doing some service to my pc soon. Might just as well do a fresh windows install including all the other extras. Hope that’ll fix things

I did a direct comparison between the two with profiler and found out that the performance gains is actually due to TSR quality and Lumen Quality in 5.1/5.2 being the equivalent of Low on High settings. So if you turn these down to Low in scalability settings in 5.4 you’ll see your frames jump right back to where they were. It’s actually a bug that the editor lumen quality settings are not being applied correctly or respected. A few bug trackers that I followed to repo and profile helped shed some light on a few really big hitters.
Unreal Engine Issues and Bug Tracker (UE-225380)
Unreal Engine Issues and Bug Tracker (TM-15457)

5.1

5.5

TSR and Lumen cost are just insane on very very simple scenes in later versions of the engine.

6 Likes

Yea, but both of your GPUs are Nvidia. I have totally the same observation as your measurements (see my UEInsight profile above). Did you compare specifically the RHIT times? They should’ve gone up like crazy between the current 5.4 and 5.3.

This was verified by multiple team members with high-end RTX GPUs.

However, another team member on an AMD wasn’t affected that bad. His UEInsight profile looks as it did before the engine update.

What surprises me the most: This is a thread with more than 20k views and almost 100 replies. Yet it goes seemingly unnoticed, almost ghosted, by Epic.

Acknowledging the issues and signalling that it’s looked into would go a long way in calming the worried devs (including myself).

3 Likes

Yeah… take a look at the 5.5 preview thread. It’s filled with comments like this trying to bring light to somewhat seemingly ignored concerns. I have linked as many ongoing issues in that thread as I can that are “essential core features and concerns” … not without a bit of backlash :rofl: but I’m trying.
Unreal Engine 5.5 Preview - General / Announcements - Epic Developer Community Forums

You are surely right, but I refrain from commenting on preview versions of the engine, as it’s most likely not getting the desired reach and impact.

However, I’m worried that dev feedback gets more and more ignored while UE turns into a sandbox for testing experimental (yet admittingly impressive) features.

But if your performance collapses by 50% from one “production-ready” UE release to another, it makes you wonder where Epic’s priorities lie these days.

3 Likes

Hi

Could you tell what exactly did you tweak in the settings for that, please?
I am doing my hobby project in my spare time and I am not 100% well informed about the UE, but I like to do my project and I was appealed with performance when I needed to move from 5.3 to 5.4 (Google Play forbids android apps being built for an older OS versions, otherwise my Android game would not pass Google Play if it was built under 5.3 version)

So… before I uninstall the old versions of Unreal I decided to put this to the test. Created a Third Person Template project in 5.1 (desktop, raytracing disabled). After shaders finished compiling I measured in standalone and then converted in-place to the next version.

I don’t have 5.2 installed, so that is absent.

Results:




Some conclusions here:

  • 5.3 was a noticeable improvement in performance over 5.1. Most of that seems to be coming from Lumen optimizations.
  • 5.3 and 5.4 are within margin of error of each other
  • 5.5p is slower, but if we look at the profiler we’ll discover nearly 1ms of that is due to 5.5 preview having a more expensive default volumetric cloud material. Lumen reflections are also more expensive, which may explain why they look so much better in 5.5p.

Bottom line is I don’t see any meaningful performance regression between versions using the starter template. I have some marketplace content downloaded so I’ll probably try in a real scene before I uninstall 5.1/5.3.

2 Likes

Just out of interest. Have you tried clearing your shader cache? Not the one from Unreal, but the one created by the driver.

I would love to see the same test done with a more demanding scene maybe with some nanite dense environment etc, I think that could produce interesting results.

So… real scene. Again, started with the third person template, this time I turned on Raytracing for Lumen + Hit Lighting for Reflections and HQ Translucent Reflections. I also enabled Virtual Texturing since this level used an RVT for the landscape material. Otherwise all settings are default.

Some things I did to make it better reflect a real modern UE project:

  • Everything that can use Nanite is set to use it (shadow method is set to VSM)
  • Set trees, boulders, and other larger foliage meshes to affect distance field lighting.
  • Grass and small vegetation are set to use contact shadows.
  • All meshes have a WPO disable distance (75 meters for trees, 50 meters for smaller foliage)
  • Unplugged PDO in the foliage materials (as this tanks Nanite perf)
  • Minor aesthetic changes (deleted their post process volume and made my own with increased saturation and local exposure, used Unreal’s built-in SkyAtmosphere instead of their included sky system)

As I recall this level is over 1 km^2, and it’s primarily foliage.

Final note: Nanite landscapes were apparently bugged in 5.1 with materials sampling RVT, which caused it to render with a glitched material. Didn’t bother trying to narrow down the issue, as I doubt it impacted performance, I only mention it in case you’re wondering why it looks messed up compared to later versions.

All measurements done in standalone process.

Results:




Conclusions…

  • Every version offered some performance gain until 5.5p, Nanite/VSM performance in particular seems to have steadily improved.
  • Can’t find any performance regressions between versions… :person_shrugging:

Performance could be better, but that is mostly because this scene was never optimized for Nanite, all the foliage is using low poly masked cards. I only did the bare minimum to get it working correctly.

Anyway… hopefully someone found this useful.

System specs for reference:
3900x
32gb ram
RTX 3070
Windows 10

Edit: I noticed an error in my original screenshot here; 5.3 (apparently) doesn’t warn you when the Nanite landscape needs to be rebuilt, and I didn’t notice. I’ve updated the screenshot so it is now a correct comparison with the rest of the versions, the timings didn’t change much (0.45ms improvement mainly from Shadow Depths with Nanite landscape enabled) so my conclusion is still the same.

7 Likes