Is the developer noticed about performance issues in EEVEE Next?
So, with the new EEVEE,everything was changed from shadows system to the new SSGI but I've noticed that new eevee struggles a lot on heavy scene and I'm pretty sure it's because of the new VSM shadows.
Also,turning on something (like "raytracing") on EEVEE Next froze blender for solid 10-15 seconds.
I hope that EEVEE Next got more optimization update on next version of Blender.
21 Comments
@jeroenbakker, thanks for your post and your efforts. The drop in performance isn't an annoyance: It's a showstopper. What's even more concerning is that the Blender Devs have been dismissive of concerns about this drop in performance: Instead of saying "We're concerned too, and are determined that EEVEE NEXT should better or at least equal EEVEE Legacy" they literally shutdown the conversation, telling us didn't need any more feedback, thank you very much. Concerns that were raised before 4.2, and remained unaddressed even as 4.2 went live! devtalk.blender.org/t/blender-4-2-eevee-next-feedback/31813/1335 I can't understate how much pain and inconvenience this has caused, which makes that last post very tone deaf. It gives the appearance (perhaps that correct appearance), they're not interested matching EEVEE Legacy performance. Thus the concern of users.
Going from 4.1.1 to 4.2, Blender's EEVEE rendering performance has dropped so badly it's now unusable for us. We use Blender to produce an animated series, and scenes that used to render in a couple of hours in 4.1.1 now render in LITERALLY DAYS on 4.2. (Check the benchmarks posted on Blender Dev). It's untenable for us.
Likewise the calls to leave EEVEE Legacy (which had been working fine alongside EEVEE NEXT until 4.4.1) intact, to cover animators in our position, had been dismissed out of hands by the Blender Devs as impossible. Yet up to and including 4.4.1 it wasn't. If it's going to take "take 1 or 2 releases to get the most annoying things settled", why not leave EEVEE Legacy there until such time EEVEE NEXT can take over, if ever.
There was a discussion on reddit discussing how EEVEE Legacy and EEVEE NEXT are very different beasts, and that giving EEVEE NEXT such a similar name was very confusing: That if they'd called EEVEE NEXT "Bubblegum" or something, I don't think we'd be even having this conversation. As in: "Bubblegum offers a great compromise between EEVEE's speed and Cycles' quality, but we won't remove EEVEE because large animators rely on its high speed." No one runs EEVEE (Legacy) for its quality. We run it for it's speed. If we all had supercomputers, we'd all be running Cycles. EEVEE NEXT takes away that speed advantage. If you only do single images or short animations, you probably won't notice or care or understand what the fuss is about. But if you do large animations (lots and lots and lots of frames with large scenes) you really feel it.
From the dev talk comments (and the 29+ upvotes for the Blender Dev who shut down the discussion!) there's an obvious schism between those who have been hurt by the drop off in performance (people doing large animations) and those who don't get what the big deal is (people not doing large animations)
From their posts, it seems the Blender Devs are in the latter camp, otherwise they'd be very alarmed: This creates the very worrying possibility that EEVEE Legacy's fast rendering is now a thing of the past, and that from now EEVEE will be much slower; essentially a faster version of Cycles.
Which means those on us who were using EEVEE for it's speed are now in a really bad place.
Our own production is suspended because of this. Should we continue, we're now stuck on 4.1.1 indefinitely, and I share OP's worries that we will never get the fast EEVEE rendering back. We obviously can't stay on 4.1.1 forever, so while we'll still use Blender for modeling and sculpting, will have to switch animation and rendering to Unreal. Which I really don't want to do, but for our purposes - large animation rendering - EEVEE NEXT and Blender 4.2 is unusable.
BF need to factor in Unreal: Rendering speed is really important. Unreal's frame rates are already better than Blender (more so since 4.2) - see that Dev thread. Although Unreal is no longer free for filmmakers, if you're earning $1M+a year a license now costs ~$3K a year(*), but if you're getting support with that, it's money well spent. If BF isn't responsive to studios - who need rendering speed (for productivity, reduced costs, faster turnarounds) - the kind of customer who falls into that category - they'll be sending those customers and their money to Unreal.
I understand EEVEE NEXT is young software, and in theory it could catch up, but my biggest concern is that the Blender Devs have been dismissive of performance concerns and haven't addressed them - despite the complaints. Shutting down those conversations in Blender Dev and Blender Artists comes across as tone deaf.
Please don't mistake this for Blender bashing. I really truly deeply love Blender and have invested a lot of time mastering it. I'm thankful for all the efforts of Ton and BF and the Blender Devs over the years. Until now it's served us very well, but EEVEE NEXT and 4.2 has been a disaster for us.
@jeroenbakker, since you tagged this proposal as incomplete, I'll make two specific requests which would address this: (1) Bring EEVEE Legacy back until such time EEVEE Next can take over, if ever, (as a former software engineer, I find the arguments made against this unconvincing) and (2) commit to EEVEE NEXT at least matching EEVEE Legacy's performance. (After all, why go backwards?) Thanks @jeroenbakker for your efforts and for replying.
Also see: blender.community/c/rightclickselect/9pjY/ blender.community/c/rightclickselect/PMw1/
@jeroenbakker, Unfortunately Blender 4.3 has not fixed the performance problems with EEVEE NEXT blenderartists.org/t/blender-4-3-released/1558823/3
Well written comment, It's such a bummer that the dev decided to close the dev talk discussion... For now I'll keep using Legacy on 4.1 because EEVEE next is just impossible to use right now! . If any developer reading this, eevee don't need to be like Cycles, It's supposed to be very fast, not SLOWER and NOISIER than Cycles . I know that you guys are working hard on this but please take your time to optimize the engine get more feedback from the community to make sure the engine is ready for studio use!
For now I don't see EEVEE Next as a good raster engine, It's just a worse version of Cycles that run as better or worse than Cycles.....
It's so sad, keep adding features but ignore serious performance problem :(
Hey @sad_mogus, It's obvious at this point the EEVEE blender devs are aware of the performance problems with EEVEE 4.2/3( @pablovazquez @jeroenbakker @ThomasDinges do you disagree?), but are ignoring them.
I think shutting down the discussion on devtalk claiming it was because of "personal fights" and because it was "now an official feature" was a bad example of this. Most people in that thread kept it professional, and it was shut down shortly after I posted benchmarks showing how badly EEVEE NEXT performed compared to EEVEE legacy : devtalk.blender.org/t/blender-4-2-eevee-next-feedback/31813/1330
As for "now an official feature", that's like Boeing saying they don't care about a defective part because the plane model is now officially in service.
The irony was on the post that shut down the devtalk thread, @ThomasDinges told us to post in RightClickSelect instead. Where we have. And we've been ignored.
In theory the addition of HDRI lighting and ray tracing(*) in EEVEE NEXT would have been nice, but not at the cost of normal 3D lighting, or performance, or shadow quality. As this Youtuber shows, you have to do all this mucking around to approach EEVEE Legacy's quality: youtube.com/watch
Some quotes:
(*) = I can fake HDRI and ray tracing using strategically-placed area lights. Yeah, having that built in would have been nice, if it wasn't at such a high performance cost. I think EEVEE NEXT is one of those cases where programmers think they have a nice idea which when implemented doesn't pan out.
Anyway, after giving up on 4.2, I tried upgrading to 4.3 but it hasn't improved, so @sad_mogus you and I and others are stuck on 4.1.
The only off-world option is to jump to Unreal 5.5. The irony is while 5.5 has incredible support for huge numbers of lights, EEVEE NEXT went the opposite direction! But Blender's strength is being able to do so much in a single pipeline, so I'd rather still do everything in Blender. But to do that, I need a fast 3D engine, which Blender 4.2/3 is now lacking.
So at this point, I think everyone can agree on this:
EEVEE NEXT is still delivering poor performance compared to EEVEE Legacy. This isn't an issue for those doing single frame or very short animations who like EEVEE NEXT's Cycle-like rendering capabilities, but it hurts those doing large animations who preferred EEVEE Legacy's Unreal-like rendering capabilities.
The Blender devs have no intention of restoring EEVEE NEXT to EEVEE Legacy levels of performance. I remember one of the devs saying EEVEE NEXT was young software and it would improve with later releases, but's that a generic statement that applies to any software! More concerning is following the Blender dev minutes (maybe I missed it?) I've seen no talk of optimization. So I conclude it's not on their radar. ( @pablovazquez @jeroenbakker @ThomasDinges Is that a correct interpretation?) Likewise ignoring us in RCS.
Which leaves an alternate solution: Bring back EEVEE Legacy, alongside EEVEE Next and Cycles, for those who need a high-performance 3d engine more like Unreal than Cycles.
I conclude the Blender devs aren't going to do this, but I'm an ex-coder and willing to put the work in. Comments? From anyone? Blender devs?
I've started brainstorming for this, not my first rodeo as I had to revert the removal of the C++ poselib in my personal builds all throughout 3.x. This removal really cheeses me, so i'll be happy to share any help I can.
It looks like the bulk of EEVEE Legacy was actually fairly cleanly cut out: projects.blender.org/blender/blender/pulls/122433/files . This seems like a massive commit at first glance, but the majority is simply throwing the entire render engine tree into the garbage, with the integrations to the scene and render systems left visible and manageable.
However, future commits muddy things greatly regarding cleanup of integration/panels/references: projects.blender.org/blender/blender/pulls/130873/files projects.blender.org/blender/blender/commit/8960285921f54c189a961d4c20022048cb529c2c projects.blender.org/blender/blender/commit/e10b0b3449788de5f51500c090ee5530652a0502 projects.blender.org/blender/blender/commit/355c7b788ec5cadbf4b25c7771653d8c6faa00f8
Those are just the first few i've seen, there are likely no less than a dozen more things that need corrected to reference both engines. The alpha build that contained both engines may have omitted this kind of linkage.
Should this be defined as the goal for this thread, providing a branch with both EEVEE engines? I'm certainly willing to sink a weekend into this sometime.
Tried 4.3, Instantly revert back to 4.1
I was frozen on 3.x for a time before writing an addon to address
poselibremoval and finally being able to update to later versions.Didn't even realize all this had gone on until testing my addon on 4.2/4.3. Like you, I'm also finding myself frozen on 4.1.
55 sec with Cycles but 5 minutes with eevee next. Damn, that's a huge difference
Blender 4.4 release and 4.5 alpha: no improvements in the least. EEVEE Next is a hard regression to EEVEE Legacy both functionally, and performance-wise. It won't be otherwise until one can open files created in 4.1 with no difference in presentation or rendering.
The developers were eager to obliterate a workflow that they didn't understand. EEVEE Next was designed with raytracing in mind and adjustments must be made to imperfectly mimic the old Legacy shadows and bloom. There's no way forward with the new engine for NPR artists like myself that used the legacy shadow+bloom+ambient occlusion system.
4.1 has not seen an update in almost 1 year now. As time goes by, they are unnecessarily making it more and more difficult to potentially re-introduce the legacy engine as a side option in order to stay up-to-date with other improvements. This is only in the last month. projects.blender.org/blender/blender/commit/894c7fa4 .
Performance is still abysmal: Fresh copies of 4.4 and 4.5 alpha both with default preferences take 45 seconds to load into Material Preview on my hardware. 4.1 takes less than 2 seconds to do the same, only to compile the default cube's shader without custom assets of any kind.
It's also rather depressing that all discussion regarding this seems to have been left behind in 2024. People should be at least critical that the Flow poster is featured as the 4.4 splash, when the movie was produced using the Legacy engine. Not a soul has noted this.
With the old engine completely abandoned on one single commit: projects.blender.org/blender/blender/pulls/122433/files , it's frustratingly within reach. If they're keen to rewrite history and have EEVEE Next take the EEVEE name in the codebase, it may actually be easier to re-implement this commit as
engines/eevee_legacyor similar.Unfortunately, it's impossible to package as an add-on, and I doubt the development team would even acknowledge a merge request for it.
Hey @breakingspell ! I am a former C++ dev with graphics experience though not Blender dev. However I am willing to contribute here if I can do it as part of a team. We've been talking about it here: blenderartists.org/t/eevee-development-updates-eevee-next/1368430/1059
We are all aware and looking in optimizations. Performance tuning options are also available in the preferences. the 10-15 seconds I asume your running NVIDIA and didn't check the
Max Shader Compilation Subprocesses( developer.blender.org/docs/release_notes/4.2/eevee/ )According to other users shadows can be easily optimized by disabling shadows on geometry that are far away. We are also looking in improvement here as well.
And yes this is the first release and we are improving it continuously. Similar to EEVEE it might take 1 or 2 releases to get the most annoying things settled.
I marked this proposal as incomplete as it is not sure or specific what is being asked to do differently feature wise.
Hello, not the original poster, but I noticed nobody has really defined a goal for this thread.
This thread has documented performance issues and incompatibilities of the new EEVEE Next system compared to EEVEE Legacy, they have not improved over time.
However, this thread has no purpose if developers are not going to worry about 1:1 feature and performance parity of EEVEE Next with EEVEE Legacy features.
The core issue is that EEVEE Next was designed with ray-tracing in mind and never should have replaced the old system.
There was a disconnect with the core purpose of EEVEE as the low-performance non-raytraced renderer, to complement Cycles (remember Workbench?). Turning the ray-tracing feature off does not restore the old shadows, as EEVEE Next also has many other changes under the hood.
It's obvious from work in the codebase that there is no going back. I'm personally awaiting a custom fork or other effort from the community in order to support the EEVEE Legacy engine moving forward.
I think it's wild how EEVEE Next was shoehorned into the LTS release and completely replaced the old version, amid the backlash it received. What is an LTS, then? EEVEE Legacy was just fine as it was to be left as a deprecated feature.
This breaks years of my toony assets and locks me into 4.1 for everything. At least i've noticed some addon developers are also providing 4.1 holdout branches.
It shouldn't have to be said to a professional team of developers, but when you develop a new software feature, you set it as the default with a fallback, and slowly deprecate the old feature.
This is just like when the (still-incomplete) Pose Asset system replaced the Pose Library system in the span of a few months. The old API library was yanked clean out of the codebase, orphaning any previous pose data. Surprisingly there wasn't hardly any noise there though, compared to this EEVEE change.
From a few hour's research on the codebase, it seems it would only take a few more hours of git reverting and linking in order to assemble a current fork with both EEVEE iterations able to coexist, just like the alpha that people have mentioned. It doesn't look like anyone else has done it yet though, so I may be underestimating the effort. It would just be another fork in the sea and not suitable for real use regardless.
And regarding performance! I could go run errands in the time my old scenes take to open in material preview in 4.2+. Thank you very much @iCameiSawiBlended for benchmarking real data that demonstrates precisely how messed up this is.
Freaking wild. Just my thoughts while I hunker down on 4.1 and keep my eyes open.
Outside this forum and devtalk,every blender youtuber seems to praise EEVEE Next like it's the polished next gen engine (100% bet that most of these youtuber never/rarely use eevee legacy)
Youtube's representation of Blender has always been a positive one. And it's true, because this software is indeed powerful, indispensably useful, and FOSS!
I'd like to draw attention to this video as well: youtube.com/watch . Not clickbait as it seems, but in fact a full proposal for a new NPR system.
It's interesting to notice how many top comments note that the "1:56" timestamp should have opened the video, where the dev talks about how they are collaborating directly with Blender to have their engine implemented.
Not a soul has noted at "1:45" where the developer points out how 4.2 "deprecated" EEVEE Legacy, and leads into his announcement of the upcoming replacement.
It wasn't deprecated, but fully removed: their engine based on Legacy was broken as well. At least they do have plans to properly define NPR in Blender.
I'm also very interested in why Blender is such a software that could freely combine several different mechanics and cloned or similar tools suddenly begins to radically cut out what works and adds something that works somehow, in my opinion it all looks like an outright miracle. While large corporate mega-giants are literally fighting for every inch of productivity, their earnings depend on it, and the developers of Blender3D decide to remove a working tool for the sake of some unknown goals by simply brutally cutting it out.
You know, I came in here to say the same damn thing. This is ridiculous. I am running a 64 gigabyte RAM computer with a 5080 in it and I can't get a consistent 5 FPS in Eevee Next with something Eevee Legacy handles BETTER.
I DO NOT SEE why Legacy has not been included as an LTS option of SOME sort, a transitional piece as Eevee Next continues development (I understand LITERALLY why, it was explained in this thread, but I don't understand why you would make such a DRASTIC CHANGE with an engine so unproven in the first place, you shut yourself out of a working solution and forced everyone to come along with you). Right now, I have switched to Cycles permanently, where I get a MORE CONSISTENT FRAME RATE. That's right. CYCLES.
It's fucking nuts. I'm normally not someone who reduces themselves to swearing when giving critique but the blase attitude about this from the development team, one I have regularly given money to support when I can, is fucking ridiculous.
I hope they fix this.
4.5 is around the corner and eevee is still unusable :/ , Another version to ignore i guess
Unfortunately, it didn't happen. Eevee Next was still crap in terms of performance. But the picture got better, that's true. But was it worth it? But they offer us to use the updated Workbench. Brilliant solution from the developers...
It does look a bit better than legacy but then you can just install SSGI version of blender 4.1 and use it without worrying about missing features and bad performance