RTX 3090 Ti: System is out of GPU and shared host memory

Support Technical Support
1

Hello.

I switched from 2070S to 3090 Ti because VRAM usage was a constant struggle. However I’m not sure if new GPU is faulty or something else is but it doen’t utilize all of its massive 24gb VRAM.

I made deliberaly heavy scene with lots of characters. But Blender either crashes (with monitors going dark for a moment) or puts out this error message.

Deleting some stuff from the scene seem to work but I think 15-16 gb of VRAM is as far as it gets.

System: i7 8700K, 32gb DDR4, m2 NVME to store the scene.

2

Welcome to BA :slight_smile:
It sounds like you might have some faulty VRAM- have you run any VRAM diagnostics?

1 Like
3

I don’t know of any way to diagnose it. Can you suggest some?

4

I’ve heard good things about MemtestG80 - here’s some info about how to use it: https://www.reddit.com/r/nvidia/comments/5a3fsp/comment/d9ehj4b/?utm_source=share&utm_medium=web2x&context=3

5

Thanks. I tried those and they show no errors. But curiously Task Manager shows only 15gb/22.5 used during the test (MULTI Memtest G80)

1 Like
6

Yeah… that’s weird, you should be using the full VRAM for that diagnostic :thinking: And that’s consistent with what you’re seeing where you’re only getting 15 GB in Blender, so it seems like 8 GB of your VRAM is just offline somehow. Who’s the manufacturer for your 3090 TI? It might be worth reaching out, assuming it’s still under warranty, and seeing if you can get a replacement or any information

7

Judging by this thread I need to have 48gb of RAM in order to fully utilize 24gb of VRAM.

This makes sense because I have 32 and I’m hitting 16gb of VRAM use in Blender and memtest.

I’ve never heard of this before. If true it should be common knowledge and something gpu manufacturers will advertise in advance.

8

I’m not sure how true this is… it might be, I’ve never personally encountered it, but if so, it should be a fairly cheap fix to go from 32 GB of RAM to 64 GB

9

Not for me, I already have 4 sticks.

Will probably have to stick to 16gb of vram for now. But at least I feel safe about GPU now. I’ll upgrade when it’s time to move to DDR5 and new CPU.

Thanks for replies by the way.

10

This may not help at all, but should be worth a shot, outside of running latest Nvidia drivers, etc.

In your BIOS, look for something called Above 4G Decoding and turn it on if its there and turned off.

This is often linked to Re-sizeable Bar, but I think your system maybe too old for that, so it’s possible Above 4G Decoding wont’ be there either, but worth a look.

You can also download GPU-Z (https://www.techpowerup.com/download/techpowerup-gpu-z/) this will give you some general info and checks of the GPU. Will help to show if things at a basic level, like amount of VRAM, is showing up.

11

GPU-Z and all other software shows 24gb though after I turned on ECC state it was reduced to 22.5

Shared GPU memory shows as 16gb in Task Manager.

Switched back to no ECC state since don’t seem like I need it and it’s not the culprit.

1 Like
12

Not sure I’d read a lot in that. Running Windows 10 with a 3080 Ti (so 12Gb VRAM) and Task Manager also shows 16GB as Shared GPU memory.

1 Like
13

But you have 32 gigs of RAM, right?

14

Yes I do, 4 x 8 GB sticks.

1 Like
15

Yeah, this must be it. I think it means GPU can move some stuff from RAM if just VRAM is not enough, but can’t exceed 16gb.

16

The Shared can’t exceed the 16GB, but the total should be the sum of dedicated + shared for general stuff. So in your case, 24GB + 16GB. Of course Cycles is likely a bit different and really just wants to keep and work on data that is only in VRAM. Even so, you should be able to push near 24GB (desktop, etc not withstanding) before Blender stops.

Did you check BIOS for Above 4G Decoding?

17

Hi Rodion,

Have you tried a proper VRAM check? Which OS are you using? For Windows you can try this:

Good luck

18

Above 4G Decoding was turned off. Just turned it on. I also checked for something called Resizeable BAR but sounds like Z370-P motherboard doesn’t support it.

19

Capture

Looks fine.

20

This is work principle most of the OSes.

Video cards can be works with lesser memory capacity too, but this is decrease your performance.