Running Video Cards in Parallel

G.A. Wells writes "Ars Technica has the scoop on a new, Alienware-developed graphics subsystem called Video Array that will let users run two PCI-Express graphics cards in parallel on special motherboards. The motherboard component was apparently developed in cooperation with Intel. Now if I could only win the lottery."

Comments

Quad-screen?

[Vrallis]

Hell, I couldn't care less about parallel processing for the video cards.

I want tri-head or quad-head video, but with at least AGP speeds. You can do it now, but only with PCI cards getting involved.

Re:Quad-screen?

[RhettLivingston]

As long as you can live with at least one of the screens not running quite as fast (maybe an informational type of screen as opposed to 3D scenery?), 3 screens ia really easy today. Almost all decent AGP cards these days support 2 screens at 1600x1200. Throw in a good PCI card and you've got 3. I've been running this way for years and it works well. Actually, the PCI card isn't shabby.

The only problem I encounter in Windows is an occasional tooltip coming up on the primary monitor instead of a secondary monitor. This is not the fault of the OS, rather the application is constraining the tooltip to be on the primary monitor by forcing it to be within the primary monitor's coordinates.

Note that Matrox's single board AGP solution does not compete with this. Using a high end NVidia for the main two screens provides too much of a performance advantage to give up for Matrox's slow cards. Matrox's cards, even though on AGP, run about like the PCI cards.

Regardless, when these systems become more available, I will be one of the first to put 2 video cards in and run 3 or 4 screens from my PCI Express system. But, though I like playing 3D games this way, I do it for the extra informational surface for programming. It greatly eases things to run your application on one screen and your development environment on all of the others so that you can see everything at once. And with 19" 1920x1440 monitors (which usually manage 1600x1200 with better focus than a 1600x1200) running around $250 a pop, its a very worthwhile investment.

Voodoo

[Eu4ria]

Didnt the early voodoo cards allow something similar to this ? I know they had a pass through from your 'normal' video card but i seem to remember the ability of running more and they would each do alternating scan lines.

Re:Voodoo

[kamelkev]

Voodoo was basically the beginning of the performance PC market, with tons of wierd options and card types.

Benchmarks for the old 3dfx V2 SLI can be seen here:

http://www4.tomshardware.com/graphic/19980204/

I was (and still am, although its in the junk pile) a 3dfx V2 owner, the performance of that card was just amazing at the time. The Voodoo and the Voodoo2 definitely changed the world of 3d gaming.

Also of interest is an API that came out much later for the 3dfx chipsets that actually let you use your 3dfx chipset (they didn't call it a GPU back in the day) as another system processor. If you were an efficient coder you could actually offload geometric and linear calculations to the card for things other than rendering. I can't seem to find the link for that though, it may be gone forever.

interesting technology

[cheese_wallet]

I think it is great that a company has the will to do something like this, even if it doesn't catch on. It's cool to try something new, instead of just hanging back and doing the tried and true.

I'll admit I haven't yet read the whole article, but even though it says that it isn't tied to any one video card, that doesn't say to me that it can have multiple disparate cards. If it is doing something along the lines of SLI, I would guess that the speeds would need to be matched between the two cards. And that would imply having two of the same card, whatever card the user chooses.

But maybe not... maybe it's the advent of asymetric multi video processing.

Re:interesting technology

[jonsmirl]

You can do this today with Chromium.

Chromium replaces your OpenGL library with one that farms the OpenGL drawing out to multiple machines. It's how display walls are built.

You can use the same technique for multiple card in the same box.

The real question

[241comp]

Is this compatible with Brook and other general-purpose GPU programming techniques? The use I see for it is this:

Imagine an openmosix cluster of dual-processor machines that run bioinformatic calculations and simulations. Lots of matrix math and such - pretty fast (and definitely a lot faster than a single researcher's machine).

Now imagine the same cluster but each machine has 2 or 4 dual-head graphics cards and each algorithm that can be created in Brook or similar is. That gives each machine up to 2 CPU's and maybe 8 GPU's that may be used for processing. The machines are clustered so a group of ~12 commodity machines (1 rack) could have 24 CPU's and 96 GPU's. Now that would be some serious computing power - and relatively cheap too (since 1-generation old dual-head cards are ~$100-$150).

By the way, does anyone know if there is any work going on to create toolkits for Octave and/or MatLab which would utilize the processing power of a GPU for matrix math or other common calculations?

Re:Man am I out of the loop.

[The_K4]

Since the PCI-Express spec defines switches (these are like P2P-bridges only they have 2 sub-buses) a mother board manufacturer could add 2 of 3 of these and get 4 PCI-Express Graphics ports (or 7 and get 8 ports) the problem is that every time you do this you have to share the total bandwidth at the highest level. Since PCI-Express does have more bandwidth the AGP 8x and 1/2 of that bandwidth is dedicated up the other 1/2 dedicated down. So the down-stream (where video cards use most of their bandwidth) is greater the AGP8x's TOTAL bandwidth. So this data path bottle next shouldn't be bad if you have 2 cards (might work well for 4 if they use the bus right).