Graphics Card with 2 (or more) GPUs?

[Clauzii]

The VooDooII was, I think, the only gaming graphics card with two/three ´processors´ onboard.

What about doing it with some of the new chips around, or at least something like the SLI-mode.
Would it be that difficult to make some circuit that made todays graphics chips capable of such things?

I think it is an tempting idea, huh???

 

[DannyBoy]

Originally posted by: Clauzii
The VooDooII was, I think, the only gaming graphics card with two/three ´processors´ onboard.

What about doing it with some of the new chips around, or at least something like the SLI-mode.
Would it be that difficult to make some circuit that made todays graphics chips capable of such things?

I think it is an tempting idea, huh???

I was lead to believe that the nVidia GeForce 256 Cards were the first cards to have their own Processor.

Im sure it was nVidia who revolutionised graphics by releasing a Video Card that removed the workload from the CPU and placing it on its on central processing unit, namely the GeForce GPU.

Correct me if I am wrong someone?

 

[Shalmanese]

The Geforce was the first one to have a PROGRAMMABLE graphics chip on board (hence the designation GPU). I believe the Rage MAXX series also had 2 chips onboard as did the mythical Voodoo 6 6000.

Basically its a question of economics. its cheaper to make 1 fast processor than 2 slower ones.

 

[S0me1X]

voodoo 6000 had 4 vsa100 (sp?)
voodoo 5500 had 2

 

[Peter]

DannyBoy, you believe in marketing bull too much.

Graphics cards started having their own brain when the "Windows Accelerator" cards first appeared, and that was in the days of Windows 3.1 and ISA VGA cards.

OK, that all was 2D ... 3D rendering was something that came much later, but after a brief venture of doing it all in CPU software, graphics chips got that capability added all over the place. That was in the early Windows 95 days.

 

[Jeff7]

Originally posted by: Clauzii
The VooDooII was, I think, the only gaming graphics card with two/three ´processors´ onboard.

What about doing it with some of the new chips around, or at least something like the SLI-mode.
Would it be that difficult to make some circuit that made todays graphics chips capable of such things?

I think it is an tempting idea, huh???

I think there was an ATi card with 2 processors on it; think it was the Rage Fury Maxx?

 

[Clauzii]

I know current GPUs are indeed very powerfull, but if the AGP bus had TWO slots, then it would almost only be a matter of rewriting drivers - I shure would have two cards then instead of one.

I don´t think it is cheaper to invent a NEW GPU insead of i.e. 2xNV25s which have been rolling out for a good deal of time now, making the cost per chip very little. Remember, when the design is finished and tested it is a matter of Robots doing the hard work

Then U could have a card with 4xNV25s or 4xR300 and some sort of multiplexing circuit splitting the rasters for the 4 chips, just like SLI on a card.

I was also thinking of the PCX2 chip which was indeed a GPU-DSP card. Originally @ 66MHz and sharing data on the PCI bus (bvadr). I could imagine a PCX-X running maybe 3-400 MHz and a capability to have 2 or more in the machine...

I know that the amount of data would be to much for the PCI, but on a 3GO-Bus or the above described Dual AGP solution in a 16x mode.

Me and some friends keep working on this.....

 

[Howard]

AGP bus? What does AGP stand for?

 

[AdvancedRobotics]

Originally posted by: Howard
AGP bus? What does AGP stand for?

AGP stands for Accelerated Graphics Port

 

[Shalmanese]

AGP 3.0 (ie AGP8x) spec now allows for more than 1 AGP slot per mobo. However, the reason why we dont commonly see 10 GPU's on a board is the same reason we dont often see 10 CPU's. Some things are just not very scalable. Especially PS and VS code.

 

[ProviaFan]

Originally posted by: Shalmanese
AGP 3.0 (ie AGP8x) spec now allows for more than 1 AGP slot per mobo. However, the reason why we dont commonly see 10 GPU's on a board is the same reason we dont often see 10 CPU's. Some things are just not very scalable. Especially PS and VS code.

With everything (possibly) gong to PCI Express in the future, including video cards, this might open up the possibility to have more than one high-performance card in a system. However, that would probably only help with multiple monitor setups, unless the dual cards functioned together like that one Voodoo card did (with SLI).

 

[DannyBoy]

Originally posted by: AdvancedRobotics

Originally posted by: Howard
AGP bus? What does AGP stand for?

AGP stands for Accelerated Graphics Port

Lol, For a second there I was thinking Advanced Graphics Port!

 

[Clauzii]

Originally posted by: Shalmanese
The Geforce was the first one to have a PROGRAMMABLE graphics chip on board (hence the designation GPU). I believe the Rage MAXX series also had 2 chips onboard as did the mythical Voodoo 6 6000.

Basically its a question of economics. its cheaper to make 1 fast processor than 2 slower ones.

But what about VooDoos & PCXs and such..
U mean they don´t count in because they are AddIns?

At least the PCX2 is very indeed programmable (with instruction cues and so on).

Correct me if I´m wrong...

 

[Shalmanese]

Nope, the Voodoo and ever other card of that era was fixed form, you could only do certain instructions on them. It was only since we got T&L that we could call them programmable.

I think your confusing processing and programmable. Anyway, the distinction is irrelevant to this discussion.

 

[Deeko]

I don't recall exactly why this is, but IIRC the AGP bus doesn't truly support multiple GPU's onboard? That was one of the reasons that the Voodoo5 was not a true AGP part, because of the dual processor nature.

Also, adding another GPU onboard would pump up your fillrates and triangle rates, sure. But then you have the issue of bandwidth, do you have enough to use THAT much fillrate?

 

[Sahakiel]

Originally posted by: Deeko

Also, adding another GPU onboard would pump up your fillrates and triangle rates, sure. But then you have the issue of bandwidth, do you have enough to use THAT much fillrate?

If you build it, they will come...

 

[Clauzii]

And if they come we´ll buy!

The Reason for me start talking about 2-3-4 GPUs is that it´s obvious becoming a problem
with the heat generated by a single chip.
Thats why I started thinking about 2-4 GPUs so each one wouldn´t get so hot

No I am not only thinking of a higher fillrate, but a 2nd GPU could also do the FSAA thing
together with some extra rendering power for GPU 1 etc. etc.

Basically it´s about the design of the chips, and of course it´s more costly to use two DIEs
on a card than a single one - unless they are not using the newest technoligy ie. 0.18u or
0.15u instead of 0.13 or even 0.09 which are comming later on.

By using ´old´ but known tech, the price will be much lower than using the newest and fanciest.
(The spaceshuttles are build like that, too - see Websters Encyclopedia 2002)

I would mean that inside a GPU some ´departments´ are working harder than others,
so I was thinking: A couple of T&L Chips, a FSAA only chip, and a controlling GPU, but made
with the DIE process that would be enough for that particulary chip.

About the interconnections between the chips: It could be a VERY fast serial protocol like the
new Opterons have - which would pave the way for a VERY scalable Graphics card.

And for the current games I think a 2xNV25 card would be faster than any GFX out there.
(Maybe not with all the ´Hollywood FX´, but as I have read, they´re not that fast in their
processing as one could wish either...)

(Me and some guys just bought 4 PCX2 cards (the Apocalypse 3D) for 6$ each and are tempted
to try the ideas we talk about - but getting info´s on the PCX2s are not that easy. But we think
we will be able to (with rewritten drivers of course) somehow piggyback 2 PCX2s together, and
yes, we are aware of the PCI bandwidth problem... so we might try some kind of AGP solution...)

All this is not to build the fastest card ever (yet anyway), but to get an idea of paralelling old
GPUs instead of throwing them away. Maybe the time spend will not pay back, but it sure is fun!

 

[jackwhitter]

the problem with building cards with multiple cpus is that silicon costs the same for the most advanced gfx chip as the least advanced.. there is a minimal cost incurred for the material you put into the design.. it might be nice considering how much R&D costs now, and how much it costs to build a new plant capable of producing .13 micron processes, but i think it is more cost effective for a company to research and make a single faster chip in the long term than multiple chips together... ati already tried it with its rage fury maxx card.

 

[grant2]

Originally posted by: Shalmanese
AGP 3.0 (ie AGP8x) spec now allows for more than 1 AGP slot per mobo. However, the reason why we dont commonly see 10 GPU's on a board is the same reason we dont often see 10 CPU's. Some things are just not very scalable. Especially PS and VS code.

Actually graphics are INCREDIBLY well suited for multi-processors. It's a simple concept- each GPU renders alternate frames. Voila! 2x as much FPS.

Of course, cheap GPUs doesn't necessarily mean a cheap card- when you have to use 2x the PCB, 2x the Fans, 2x the memory, AND design/add extra chips to "traffic control" the alternating frames, then you lose a lot of economic advantage. Plus, these days it's a lot sexier to create a new, faster GPU than kludge together 2 old ones just to have the enthusiast community make fun of your 'inelegant' solution.

 

[ProviaFan]

Originally posted by: grant2

Originally posted by: Shalmanese
AGP 3.0 (ie AGP8x) spec now allows for more than 1 AGP slot per mobo. However, the reason why we dont commonly see 10 GPU's on a board is the same reason we dont often see 10 CPU's. Some things are just not very scalable. Especially PS and VS code.

Actually graphics are INCREDIBLY well suited for multi-processors. It's a simple concept- each GPU renders alternate frames. Voila! 2x as much FPS.

Of course, cheap GPUs doesn't necessarily mean a cheap card- when you have to use 2x the PCB, 2x the Fans, 2x the memory, AND design/add extra chips to "traffic control" the alternating frames, then you lose a lot of economic advantage. Plus, these days it's a lot sexier to create a new, faster GPU than kludge together 2 old ones just to have the enthusiast community make fun of your 'inelegant' solution.

Render two frames simultaneously, and you might increase apparent latency in fast paced FPS games, though, unless you can render two frames in less time than a single GPU card could render one frame.

 

[0roo0roo]

from what i remember the v2 had separate chips for the texture units right? not exactly gpu's

 

[KillaBong]

My rage maxx fury rendered alternate frames from what I understood. Heh, I still want a voodoo 5 6000 (128mb and 4 proccesors).

 

[Amorphus]

ok, some things you guys are forgetting : the AGP port is not actually a "bus", per se - the PCI bus is a bus, because they're all on the same data channel (something like that. the communicate within a single channel or something...)

the AGP port has its own direct line to the northbridge, and thus, to the CPU. the thing about two AGP ports, then - if we want multiple AGP ports, we're going to have to increase the amount of bandwidth that the northbridge is going to handle by almost 25%, and we'll also need new northbridge design.

secondly, workstation graphics cards already have dual GPUs...

I saw a 3DLabs card (Wildcat) that had one... here we go:
http://www.3dlabs.com/product/wildcat4/images/wc4_7210.jpg
Wildcat4 7210 is the model, I think (might just be Wildcat 7210)

 

[Clauzii]

Originally posted by: 0roo0roo
from what i remember the v2 had separate chips for the texture units right? not exactly gpu's

That´s right, but together with the third chip on VooDoo2-boards working as a GC, Graphics Chipset,
they were. Remember, the VooDoo Chips started as a one chip solution.

--------------------------------

So if we took todays mainstream chips (NV2x´s or R3xx´s or even not-so-mainstream Parhelia´s) and designed a graphics board with two or more of these, drawing alternate frames (drawing alternate scanlines would indeed
demand a new GPU design, i think) and smack on 2x128/256MB (one block for each GPU) it would be a killer!

While a GeeFX board will cost around 4-500$, the NV2x chips cost around 50$/each. So I don´t think it
would be so Xpensive to build in the end (besides from the AlternateFrameShifter chip, but once designed
(by ATI?) it could be adapted to a lot of GPUs)....

I also know the 3DLabs cards (actually, I love them!) but they are non-consumer cards build for CAD and
the like. Not saying they can´t be used for games (OpenGL anyway).

No I´m not ´forgetting´ that the AGP is a PORT and not a bus, being bounded to the motherboard
Nothbridge.
If what was said earlier about AGP 3.0 specs (supporting another AGP port) is right, it IS
indeed only a matter of time before we´ll see multicard consumer solutions again. Maybe not working as
a doubled FPS GPU but more like a ´SurroundView_NoFPS_Cut´ Parhelia-like tingy

A dual AGP system could also be made by dividing a AGPx8(or 16) Port to two AGPx4(or8) ports, using time
slicing, thereby enabling 2 cards at the same time, sharing the IRQ, but being (driver)-recognized as a
DualAGPort capable. I don´t know the actual transfer protocol of the AGP-Northbridge, but I can´t see
why it shouldn´t be possible....

Imagine having TWO Parhelias (SIX screens): Thats surround gaming

...VooDoo5 6000... sounds like a dream....

 

[Clauzii]

Originally posted by: Deeko
I don't recall exactly why this is, but IIRC the AGP bus doesn't truly support multiple GPU's onboard? That was one of the reasons that the Voodoo5 was not a true AGP part, because of the dual processor nature.

Also, adding another GPU onboard would pump up your fillrates and triangle rates, sure. But then you have the issue of bandwidth, do you have enough to use THAT much fillrate?

It might not be a true AGP part, but it works, doesn´t it?

Put more memory on the Cards....