Nvidia for compute, Aspeed to drive the monitor.

balnazzar

Active Member
Mar 6, 2019
205
25
28
Hi folks.

I have a supermicro board with IPMI, including an Aspeed AST2500 video card.
Now I was wondering if I can use it to drive the monitor, leaving the Nvidia card(s) just for compute. Does anyone of you attempted this? Should one use the Nvidia headless drivers? Thanks.
 

BoredSysadmin

Active Member
Mar 2, 2019
564
172
43
Am I missing some context here?
If you would like to run this server headless you don't need a video card, use IPMI to manage it remotely when needed.
If you DO need to connect a monitor to it to use all the time, I'd get some cheap/low power GPU for that.
 

balnazzar

Active Member
Mar 6, 2019
205
25
28
I didn't clarify the context. Running a graphical interface on a video card consumes vram upon it. I do have four video cards, and when running models in parallel, ideally the cards should have the same amount of free memory. Even better, all of their memory should be free.
On the other hand, it's known that if you want to use the regular nvidia drivers on linux, you got to run an x server.

So, a functional solution for me would be: run the x-server with the Aspeed 2500 (I don't need high graphics perf, but I still need a graphical interface), AND run the four Turing cards with the nvidia-headless drivers.
That is, IF this is possible.
 

RageBone

Active Member
Jul 11, 2017
361
96
28
that should be possible.
Windows likes to switch away from the BMC-GPU leaving you with a blackscreen but since you are running linux, that shouldn't be an issue.
Just Try it?
 
  • Like
Reactions: balnazzar

balnazzar

Active Member
Mar 6, 2019
205
25
28
that should be possible.
Windows likes to switch away from the BMC-GPU leaving you with a blackscreen but since you are running linux, that shouldn't be an issue.
Just Try it?
Yep, but I think it won't be so easy, particularly on Nvidia's side...
 

larrysb

Active Member
Nov 7, 2018
107
45
28
The current nvidia drivers can run headless.

Typically I run a docker image(s) with the framework(s) intended on an ubuntu server with python notebooks remotely to one or more mac or windows laptops on the LAN.

I've tried with the Asus Workstation IPMI/AST2400 motherboard in vga mode and it didn't seem to work all that well. The AST2500 might be better.

On the Turing cards, with NVLink being limited to 2 cards and the cooling issues of the fan-style cards, I'm down to two RTX Titans per workstation, even with the 4-way SLI capable motherboards I'm using. (PEX switches for 4x x16pcie slots and Xeon E5-v4 CPU). But have been using distributed training over 25gb ethernet between instances.

I'm limited mostly by the amount of power available from the wall outlets. o_O
 

balnazzar

Active Member
Mar 6, 2019
205
25
28
The current nvidia drivers can run headless.
You mean the regular drivers or the server drivers (the ones marked with 'server' in the package name)?

I've tried with the Asus Workstation IPMI/AST2400 motherboard in vga mode and it didn't seem to work all that well.
Mmmhh the video card should be the same. I use it with a vga --> hdmi converter, and it's indistinguishable from any other video card at 1080p. Try it.

May I ask you if suspend/hibernation *works* when the Aspeed VGA card is active?
To make sleep/hibernation work on my supermicro, I have to disable it altogether by the dedicated jumper. Otherwise it prevents linux from acpi s3/s4...