One display with multiple GPUs

[Neanderphil]

Hi

When running two videos cards in a Mac Pro, or any other desktop for that matter, is it the card that is connected to the display that is responsible for any hardware acceleration?

For example, take a Mac Pro with a GTX 285 and a GT 120. A display is connected to the Mac Pro via the mini display port on the GT 120 and the GTX 285 is connected to nothing. Will any GPU intensive software, for example a game, be accelerated only by the GT 120 or is possible to have the GTX 285 undertake any open GL or direct X tasks and the GT 120 just drive the display.

I'm aware that SLI is not supported on the Mac Pros. I have tried googling but failed to come up with anything conclusive.

Thanks

 

[zmttoxics]

The card that handles the screen handles the acceleration.

 

[Cuda]

The card that handles the screen handles the acceleration.

That was not always true, may not be true now, however I'd say for most end users it's too hard to setup, so it's as good as true.

At one time you could launch your 3D app on a screen connected to the card you wanted to Render on, then move the window( Had to be in Windowed Mode) to a display connected to another card(GPU), and it would still Render on the Card(GPU) it was launched on.

It's been a long time (10.3), the last time I tried it, so things may have changed.

I'm sure it could still be setup, in some way, but I don't think any Documentation for it is public.

I assume, that the point would be to get the Power of the 285GTX to the LED Display. You'd have the overhead of Data going to the 285->Coming from it->Going to the 120, over the same Bus, so each card should be in 8x PCIE slot(Mode).

The OS X Window Server sucked a bunch of CPU last time I tried it, so that's something to overcome, too.

Stretching a 3D window a cross 3 Displays still worked, last time I checked, that does require the same thing, so it "Should work", but you would need a display connected/Detected, on the card(GPU) you wanted to Render on.

As far as I know, the MacBook Pro with 9400m and 9600mGT does some sort of this same thing, but I could be wrong.

 

[zmttoxics]

That was not always true, may not be true now, however I'd say for most end users it's too hard to setup, so it's as good as true.

At one time you could launch your 3D app on a screen connected to the card you wanted to Render on, then move the window( Had to be in Windowed Mode) to a display connected to another display, and it would still Render on the Card(GPU) it was launched on.

It's been a long time (10.3), the last time I tried it, so things may have changed.

I'm sure it could still be setup, in some way, but I don't think any Documentation for it is public.

I assume, that the point would be to get the Power of the 285GTX to the LED Display. You'd have the overhead of Data going to the 285->Coming from it->Going to the 120, over the same Bus, so each card should be in 8x PCIE slot(Mode).

The OS X Window Server sucked a bunch of CPU last time I tried it, so that's something to overcome, too.

Streaching a 3D window a cross 3 Displays still worked, last time I checked, that does require the same thing, so it "Should work", but you would need a display connected/Detected, on the card(GPU) you wanted to Render on.

As far as I know, the MacBook Pro with 9400m and 9600mGT does some sort of this same thing, but I could be wrong.

I shouldn't have used the word acceleration. But the screen rendered, comes from the card it is attached to. To render the screen, the data to sent to the card connected to the screen either cpu calculated or to be calculated by the gpu, which is stored in the buffer. So, the screen gets its image from the buffer on the card its attached too.

In the MBP with 2 videocards, only 1 card runs at time. The window server is reloaded when the cards are changed (the reason they make you log out and log back in after switching graphics modes on them).

If there is some confusion there, I apologize, I am on my 4th coffee and still tired.

 

[Cuda]

In the MBP with 2 videocards, only 1 card runs at time. The window server is reloaded when the cards are changed (the reason they make you log out and log back in after switching graphics modes on them

That's correct, I remember reading that and had forgotten, I was wrong on that point.

It does lead me to wonder how the "Switching" happens, from the point of view of the Connection to the Display to each of the 9400 and 9600? Sorry, I don't know how to ask the question I'm asking.

 

[zmttoxics]

That's correct, I remember reading that and had forgotten, I was wrong on that point.

It does lead me to wonder how the "Switching" happens, from the point of view of the Connection to the Display to each of the 9400 and 9600? Sorry, I don't know how to ask the question I'm asking.

That can be handled with a simple electronics multiplexer circuit or something of the likes. Apple wasn't very specific on how they did do it though.

 

[Cuda]

Will any GPU intensive software, for example a game, be accelerated only by the GT 120 or is possible to have the GTX 285 undertake any open GL or direct X tasks and the GT 120 just drive the display.

I just checked, under 10.5.7 PPC, it can still be done, however, not really the way you wanted to do it, you are still going to need one display connected to each card, or at least a "Ghost Display" on the 285GTX(But that's not a real option, unless you want to Screen Share you own IP and move the 3D app window to the Real Display.)

On my test, the Window Server still sucked up 70% of the CPU( SP G5 1.8Ghz ). Not leaving too much to run the 3D app( JKA ), but it did run.

I'm not sure why the Window Server easts up the CPU time.

My test setup was G5 with 5200FX AGP and R7000 PCI. I opened the game on the 5200FX in windowed mode and moved the window to the R7000.

I did check the Driver Info and make sure JKA was still running on the Nvidia card, and it was.

I'm hoping the High CPU usage is due to the AGP->PCI Data flow( limited bandwidth on PCI, and I don't know if AGP can send Data the way PCIE can?)

So, it may work better in a MacPro with two 8x PCIE cards.

The real question is, can a way be found to make it work, and be FAST/Easy/Full Screen.