Best Practices For Building A Multiple GPU System

[FrankPooleFloating]

Oh yeah.. Pssssst... You can still upgrade to Win10Pro for nada.. $0...

[smicha]

Oh yeah.. Pssssst... You can still upgrade to Win10Pro for nada.. $0...

Yes, sometimes it works

[smicha]

I am on a UEFI and GPT win10 for 1 month on a Supermicro X11 platform and scalable xeon 4110 - it rocks

PS. Once you get into Supermicro world there is no way going back :

Hi Smicha,
This is a mega Noob question, but do you run Windows or Linux with the Supermicro?

Both - win7, win10 and linux (various). Win7 - max 7 GPUs, 11 GB VRAM, win10 - 11 GPUs and 9 GB of VRAM, linux - 11 GPUs and 11 GB of VRAM.

[Notiusweb]

I am on a UEFI and GPT win10 for 1 month on a Supermicro X11 platform and scalable xeon 4110 - it rocks

PS. Once you get into Supermicro world there is no way going back :

Hi Smicha,
This is a mega Noob question, but do you run Windows or Linux with the Supermicro?

Both - win7, win10 and linux (various). Win7 - max 7 GPUs, 11 GB VRAM, win10 - 11 GPUs and 9 GB of VRAM, linux - 11 GPUs and 11 GB of VRAM.

I am sorry if I am new to this conversation and it has already been discussed a million times, but is it known why the max GB of VRAM on Win 10 is 9?
Is it the amount of GPU, or would that also be the case with 1?

[smicha]

We assume that this is due to DX12. And it is valid across any GPU count. Usually it is around 16% of vram that is taken by system.

[Notiusweb]

We assume that this is due to DX12. And it is valid across any GPU count. Usually it is around 16% of vram that is taken by system.

Mmm.....
I guess Linux really is the champion!

[smicha]

We assume that this is due to DX12. And it is valid across any GPU count. Usually it is around 16% of vram that is taken by system.

Mmm.....
I guess Linux really is the champion!

Yes

[Notiusweb]

Did you see the benchmark score posted for 17x 1080Ti @4000+ s/px.....
Is that you Tutor!?

[smicha]

This is Chris Franklin in x9drx

[Hatsize7]

Hi guys,

I need your help. I built a 12GPU system with a H110 Pro BTC+ motherboard (I bought the rig from an ex-miner) with PCI-E risers, etc, it runs smooth, Octanebench is around 1770.
There is one issue though. On medium/heavy scenes it gives Cuda error 2: out of memory (failed to allocate device array). Other slaves with GPUs connected directly to motherboards work fine on these same scenes.

I'm running Windows 10 Pro, 32gb ram, cards are GTX1080 and GTX1070Ti's, Nvidia drivers are 417.35 and sending the scene to the slaves with the C4D R16 plugin, octane version is 4.01.1.

GPUs are not out of ram, but out of core memory is enabled, just in case.
In abstrax's post here: viewtopic.php?f=24&t=69639 it was noted that it was a driver issue and had been fixed with Nvidia drivers 417.22, but I'm running Windows 10, not 7/8 and a newer driver than 417.22 and still experience the issue.

Did you guys have this kind of problem with multiple GPU rigs? Any solutions or ideas what to do (other than trying to optimize the scene)? Thank you in advance!