Archives Discussions

[tgm]

I have one HD4870 and one Firestream 9250. The system is Ubuntu 8.04.

I just need HD4870 to connect to the monitor and Firestream 9250 only used to compute. How to configure the system (xorg.conf)?

[kos]

This is realy important theme, but I don't know nothing about that, and there are no moderators on the forum, ot they just can't answer on this.

[udeepta]

You can think of 9250 as any other video card. While setting up the system, the only thing you need to be remember is that the 9250 does not need to be connected to a monitor to be active, while Radeon cards do need to be hooked up to a monitor.

So, to answer your question, configure xorg.conf as you would for any two video cards. No need to do anything special. If you have set up everthing correctly, the CAL sample FindNumDevices should show two devices.

Btw, FireStream cards can live alone without another video card -- they can drive 2D video without much stress on their compute ability. But that is not to say that you should not have a 4870 along with a 9250 -- you now have two powerful GPUs! 🙂

When using CAL, there are APIs to specify which card you want to run your compute programs on. When using Brook+, you set the BRT_ADAPTER environment variable to 0 or to 1 to direct your program to run on the first or the second GPU. We are planning better multi-GPU support for Brook+ 1.4.

[tgm]

I'm afraid that it is not my expected answer. Note that there is a statement "Ensure a monitor is plugged into the graphics card. A monitor is required for each card in a system with multiple cards." I only have one monitor.

[udeepta]

Let me rephrase:

1. Imagine you have two 4870s. The xorg.conf file that you will use in that case is the same as the file you will use for 4870+9250 configuration. I cannot test it for you right now; but i am sure the driver help files and google will turn up something for you.

2. Trust me, 9250 will work without being connected to a monitor -- remember to use FireStream drivers. 4870 will not work if it is not plugged to a monitor.

3. If the setup is giving you trouble, take out the 4870 and use the 9250 to do both jobs -- driving the monitor and doing compute.

Hope this helps.

[kos]

So I can not use your cards to buils gpu cluster or to make a desktop supercomputer and just have to pay for amd to allow me to do that ? Oh, please, let me use your cards in my application. nVidia sells it's pro cards with mora RAM amount, but yours are just the same as any other gaming solution (I don't need 30 bit color for computation). Are you kidding ?

[tgm]

That would be interesting.

[udeepta]

Kos, what is preventing you from creating a gpu cluster? Can you be more specific and give us some details so we can try to help you?

On another note, our pro cards have a much better performance/cost and performance/power ratio than others out there; but ultimately it is for our users to decide what they like.

Originally posted by: kos So I can not use your cards to buils gpu cluster or to make a desktop supercomputer and just have to pay for amd to allow me to do that ? Oh, please, let me use your cards in my application. nVidia sells it's pro cards with mora RAM amount, but yours are just the same as any other gaming solution (I don't need 30 bit color for computation). Are you kidding ?

[kos]

So, I have currently 3 machines : 1 master node, 2 slaves with different AMD's gpus. Now I'm only configuring cluster software and considering the ability to write my own. Machines in cluster have to work (compute) with no display. And now I'm hearing that mainstream card's couldn't work(compute) with no display attached to it. Is this right ? So I can't understand one thing : I need monitor to be connected all the time or just to setup the driver, like in CUDA ?

[udeepta]

Kos,

There are two experiments you can do to figure out. Sorry I do not have a setup near me right now, so I cannot do it for you.

1. http://forums.amd.com/devforum/messageview.cfm?catid=328&threadid=98907&enterthread=y

Try this: Boot your machine, start X Windows, remove the monitor, and run a stream app remotely on the machine. AFAIK, it should run. Now try again but remove monitor before you start.

2. Take a dual GPU machine. Connect both GPUs to monitors. Boot your machine, start X Windows, remove the monitor from the second GPU, and run a stream app on the second monitor (use BRT_ADAPTER). AFAIK, it should run. Now try again but remove monitor before you start.

[kos]

So I need monitor only to setup hardware and driver ?

[tgm]

That's definitely unacceptable! Can you imagine how to setup a cluster with thousands of dual GPU nodes?

Even in a desktop machine with dual GPU cards, what if I have only one monitor?

Originally posted by: udeepta@amd Kos,

2. Take a dual GPU machine. Connect both GPUs to monitors. Boot your machine, start X Windows, remove the monitor from the second GPU, and run a stream app on the second monitor (use BRT_ADAPTER). AFAIK, it should run. Now try again but remove monitor before you start.

 

[michael_chu]

There are a few ways around this today:

- You can use a dummy VGA "adapter". This is an "adapter" that grounds the RGB lines from their respective grounds with 75ohm resistors.

http://soerennielsen.dk/mod/VGAdummy/index_en.php

We have done this in our labs and for demos. Easiest way we have done it is to get a DVI-to-VGA converter (I forget if one is included with the Radeons) and insert 3 75ohm resistors in there.

This will make the card think there is a VGA monitor attached.

- You can use a FireStream card which already allows you to use the card without a monitor attached.


There are other potential solutions being discussed internally but I can't comment on their schedules just yet.

Michael.

[kos]

That sounds better. I cann't imagine big cluster administator who runs around failing from time to time servers with monitor.

[JonBoyd]

I've run across this problem before and never had a good solution. I like the VGA adapter with the 3-75 ohm resisters tied to gnd. Wish I'd thought of that.

Does the card make an assumption about the resolution at that point or does it matter?

Edit:Removed Advertising from the post

[michael_chu]

It'll assume (incorrectly :-)) that you have a 640x480 standard VGA monitor attached.

Of course, not that it matters though since you won't be able to see any pixels its pushing out that port. 🙂

Michael.

[LenAlox]

I am trying to figure out how to tell Brook+ to use device 0 or device 1 (Both firestream 9170 boards) in Linux CentOS 5.1. I have seen BRT_ADAPTER is what you change, but from what I understand it is an eviroment variable and I don't know how that applies to a Linux platform.

I used export BRT_ADAPTER=1 or 0 in my .bashrc file, will this tell the brook code to go to one device or the other? Also, is there a way to tell what is running on a card in Linux?

Thank you.

[thesquiff]

So do 4870X2 cards need to monitors to be hooked up to use both GPUs?

[MicahVillmow]

If the X2 is in crossfire mode then only one monitor is required to have both GPU's visible to CAL.

[ryta1203]

Originally posted by: MicahVillmow If the X2 is in crossfire mode then only one monitor is required to have both GPU's visible to CAL.

Yes, but then won't CAL only recognize 1 device, so wouldn't that defeat the purpose?

Ok, it looks like CAL sees 2 devices for a 2 Card CFX setup with CFX enabled and only 1 with CFX disabled. It seemed from previous documentation that we were suppose to disable CFX, seems confusing.