Help with "Persistent Data"

Support Technical Support
1

Hello guys, first post here.

I’m very, very “new” to blender (and everything around this. I have backgrounds in civil engineer/development, so i’m used to modeling softwares like Revit).

We just finished a project that is a VR experience (Panoramic 360º video / 8K resolution basically), where we got another company to create the scenes for us (we’ve provided the geometric model for them to adapt into blender) - but i’m the one tasked with the render/export of the frames. It’s about 9000 frames.

I Have some OK computers (14 at total, varying from rtx2080ti + 64gb ram all the way to 5070ti + 128gb ram), that we use to process geodata mainly (generating point clouds, ortho, etc…). Well in this little time i’ve been working with blender, it seens to me that those PCs are underspec’ed, even tho they can be considered almost “super-computers” in the civil field.

BUT, my real problem lies here: EVERY rendering that I’ve tried first, had the same pattern:
first frame has an ok rendering time (6min~20min, varying with the PC)
second frame ALWAYS slows down (like, 20~30~40x the time of the first one)

After some research, i’ve figured out that what is doing this, is the “persistent data” option in the performance settings of the render. If I turn it off, all renders have a steady, constant frame export rate. But when I talked to the guys that we outsourced the project (bc every project was coming with this on), they said this slow down does not happen with them, even tho some PCs are worse than my worser one (and also pcs in the middle of my range are exporting with this) - How can my 14 computers (I tested the SAME SCENE in all of them, every single one shows this pattern) be plagued by this, while none of theirs are? Is there something i’m doing wrong to use persistent data?

I have optix selected as my cycles render device (the CPU IS NOT SELECTED HERE), backend VULKAN with the gpu as main device, everything setted to GPU in the render options, video sequencer memory cache setted to 32gb/16gb (1/4th of the pc total ram), and NOTHING solves this.

I’ve read that this option should increase render speed absurdly, and should always be used when possible, but I simply can’t use it ANYWHERE. And what baffles me more is that (theoretically) worser PCs are running with it and mines simply don’t!

Anyone can help with this?

2

I don’t have in-depth knowledge on this problem, but this sounds like the behavior I would expect when running out of VRAM and thrashing the memory on each rendered frame.

I recommend gathering more data by monitoring your CPU and GPU while rendering (e.g. in a program like MSI Afterburner). Check to see if:

  • The GPU you expect is being used (and if it’s maxing out)
  • Your GPU’s VRAM is running out
  • You see CPU usage spiking or general system memory running out (this could be a sign that the render is not using the GPU as you expect)

If you don’t see any irregularities when monitoring, then I guess you’ll have to narrow down what differences there are between your setup and your outsourcer’s setup.

3

yeah that’s exactly what I think its happening, running out of VRAM and laying off the work to the cpu… But how can this happens on my GPUs with more vram than theirs, and not with them? it doesn’t make sense at all to me!

I’ll run a quick test and comeback with prints of the usages/spikes to narrow this down

4

8K frame itself may take considerable amount of VRAM, try rendering with smaller tiles if possible to lower VRAM usage.

As far as I know, this is not really what would happen. It should still be rendering on GPU, only RAM is used instead of VRAM and that’s why the slowdown(while data is moved between RAM and VRAM). In any case if persistent data is a problem, why use it?

5

image
image2481×675 284 KB

In this image you can see exactly what i talked about! 1min 13sec for first frame, 30min+ for second one.

nvidia-smi indeed shows an almost topped vram (not exactly full, but its almost running out of it, now it’s almost static around 11322MiB).

But I still can’t make sense of how a weaker GPU don’t stutter on this like ALL of mine 14 does

6

Oh, I see, you are using tiles, that’s that then…

Maybe there is a bug in GPU drivers… Try updating or even older versions.

Yep. That’s running out of it. Some is probably reserved by the OS/other software.

7

yeah that sounds like a bug, try a different blender version, even an experimental / daily build… Try without Vulkan… I think you probably need to click random buttons in order to find where is the issue.

Basically persistent data stores the textures and maybe static geometry in memory so it doesn’t have to process it for every frame. The result is that you are just doing the same thing than frame 1 but faster…

But it’s hard to debug without having access to the computer and try some stuff …

8

I mean, i’m not using it to do the actual “production” exports - i’ve done all of that without it on. I’m reexporting some scenes now for a new version with some adjustments, and I got less time, so I’m looking at everything that can speed up my export (without compromising quality in any way).

But in reality, the real reason i’m asking/searching this is because it does not make ANY sense for me. I simply can’t understand why this happens and it’s buggin me for a lot of days now…

9

It’s running out of VRAM. That makes sense. Persistent data increases memory usage.

It’s not about some undefined “goodness” of a PC but about the amount VRAM. RTX3090 for example has more VRAM than RTX5070TI, so… what amounts of VRAM are they using to render succesfully without slowdowns?

10

For example, they can run it without a single slowdown in a 3060 with the same 12GB VRAM…

11

Well, I’ve changed from vulkan to openGL and… Voilá, all frames are rendering faster than before, with persistent data on. (aprox. 20% increased speed)

The VRAM shows the exactly same behavior, almost topped out, but now it’s not stuttering a single bit.

As I said, this still does not make any sense to me

12

You could try reporting this as a bug if your GPU drivers are up to date. It makes sense that Vulkan backend might have some bugs.

13

Where can I do this report? Thanks for the help tho! helped me narrow down the problem.

FYI i’ve tested with all drivers (studio, not game ready) from version 32.0.15.8129 (05/09/2025 - the oldest I got in my server) till the latest released version 595.79 (10/03/2026) - I’ve had the same problem with every single one of them

14

In blender’s help menu, report a bug !

Glad you find the issue !

15

Just Help → Report a Bug from Blender. It’s a bit of work to be useful to figure out a way a developer can reproduce the issue without your production files, but it’s very helpful.

16

Alright, i’m going to report it and try to be crystal clear so they can at least try to reproduce/debug this. Probably optix+vulkan+highres and other stuff working in unison to create this kind of bug, but i’ll try my best.

Thank you all for the help, you guys saved LOTS of sleep hours for me. Really really appreciated. (Also what a nice and helpful community you guys got here, i never dreamed that i could solve this, THIS QUICKLY, in a forum nowadays… remembered me of the good old time of gamefaqs and the likes… Keep up the nice work guys)

17

Persistent Data can help in some scenes, but it doesn’t always fix the real bottleneck. Sometimes the slowdown is textures, VRAM, denoising, samples, or scene setup instead.

I made a small Blender add-on called RenderMind for this exact kind of guessing problem. It analyzes a Cycles scene and shows likely bottlenecks, then gives suggestions you can apply and revert.
https://rendermind.gumroad.com/l/rendermind

It might be useful if you want to see whether Persistent Data is actually the main issue or if something else is slowing the render down.