Multi-GPU rendering

[Jax Omen]

Obviously this would have to be worked out via a comprehensive API, but it seems to me that there are two "main" factors to graphics rendering right now.

Shaders/vertex/geometry/etc

Textures/AA/other memory-intensive options.





Why is there no SLI/Xfire mode where you pair up, say, a GPU with lots of high-speed memory (and memory bandwidth) and a GPU with tons of geometry processing, and have them each render each frame, working in unison?

Is it an issue of keeping the two GPU's in sync? Latency of the SLI/Xfire bridge?

I'm just curious, it seems like a better approach than having multiple cards do all the work on half as many frames (AFR).

 

[Crusty]

I would guess it's mostly because they do not want to have to develop and maintain two totally different hardware architectures and drivers.

What about the people that don't have the money or the need to run SLI but still want the latest features from the GFX cards like new DX support, or built in PPU's on the cards?

 

[BrownTown]

Multiple GPU systems are a low volume item, it would makes absolutely no sense to develop 2 entirely different desings (in addition to the single GPU design). Also it is highely unlikely that there would be ANY improvement in graphics speed. Graphics cards already have dozens of copies of the entire graphics pipeline, there really is no advantage to putting half the stuff on one card and half on the other. In fact I would expect a huge DECREASE in speed due to this scheme since you would have to send all the information across the PCI-e instead of on chip busses.

 

[BFG10K]

Both vendors stick with AFR whenever possible as it's the only mode that doubles vertex performance. Cards working in unison on the same frame need to divide the frame at the pixel level so you don?t get the same kind of scaling as you do with AFR.

That and your proposal would likely slow things down as the delay from accessing memory on the other card would far outweigh any potential bandwidth advantages,.

Anyway, vendors often use the best memory available in high end parts, so there?s no need for a dedicated memory card.

 

[Throckmorton]

How about putting two GPUs on one board like in the old 3dfx days and having them share memory, etc, without any SLI involved?

 

[lyssword]

Originally posted by: Throckmorton
How about putting two GPUs on one board like in the old 3dfx days and having them share memory, etc, without any SLI involved?

As far as I know, 3dfx did use SLI. In fact NVIDIA bought 3dfx and it is the 3dfx people in nvidia who developed SLI. I'm not 100% sure on that though

 

[Crusty]

3dfx used SLI, who knows if it's the same technology or if they just share a name though. In fact, I used to run 2x Voodoo 2000 in SLI!

 

[firewolfsm]

A better solution would be to treat two dies as one processor, like the Core Quads, that's what I expected out of SLI, I'm not sure why they don't do it.

 

[Nathelion]

Originally posted by: firewolfsm
A better solution would be to treat two dies as one processor, like the Core Quads, that's what I expected out of SLI, I'm not sure why they don't do it.

I'm purely guessing here, but I could imagine

1) It would be far too complex to be worth it considering how few multi-gpu solutions are sold

2) Whatever bus architecture they use does not lend itself to being shared in this way (sort of like AMDs hypertransport makes gluing two X2s together difficult)

3) It would require a new memory controller

with some overlap between the options

 

[BFG10K]

How about putting two GPUs on one board like in the old 3dfx days and having them share memory, etc, without any SLI involved?

There was SLI involved, but it was SFR (each board worked on the same frame) done at the hardware level. That's why it was more compatible than today?s solutions, which mainly do AFR at the driver (software) level.

3dfx used SLI, who knows if it's the same technology or if they just share a name though.

They only share the name. 3dfx's method was Scan Line Interleaving.

The original version alternated lines between boards while the VSA-100 boards (e.g. Voodoo 5) could do variable scanlines on each (up to 128 in each block IIRC).

The closest we have to that now is ATi's Super-Tiling which splits the frame into small alternating blocks at the hardware level, and I think it's a damn shame they scrapped it. It won't always scale like AFR but it should be more compatible and there's automatic load balancing at the hardware level.