AFR: Preferred SLI Rendering Mode

[Dave Baumann]

ftp://download.nvidia.com/developer/presentations/2004/GPU_Jackpot/SLI_and_Stereo.pdf

This is a recent (2-Nov-2004) developer document that discusses how SLI works and also looks that what to avoid for SLI use. They are suggesting that AFR is the preferred rendering mode rather than the Split Screen as the Split Screen has a greater communication overhead and scales less.

 

[digitalwanderer]

Any benchmarks?

 

[Cryect]

What a surprise they want people to use a method that almost doubles the framerate. Wonder does the slave video card only have a single framebuffer? Seems it doesn't need two except in SFR while its still sending the last frame I guess.

 

[pcchen]

I still don't like AFR based approaches. However, AFR are easier as there is almost no need to communicate between two cards. It automatically solves many problems, including load balancing, geometry scaling problem, and render-to-texture problem. And it does indeed make games look fluid, although the response time does not improve (most people don't need much of that either).

 

[Pete]

If AFR is single-buffered, does that mean there isn't be a response time hit compared to a typical double-buffered card?

 

[digitalwanderer]

Just to make sure, AFR=alternate frame rendering...right? :|

 

[Rys]

Just to make sure, AFR=alternate frame rendering...right? :|

Correct The cards take it in turn to draw alternate frames.

 

[kemosabe]

Just to make sure, AFR=alternate frame rendering...right? :|

As in something ATI described and implemented five years ago - another Nvidia "innovation".

 

[digitalwanderer]

Just like ATi's old MAXX hardware used to do it, right? Voodoo did the interleaving-whatevermajiggy, right?

 

[pcchen]

If AFR is single-buffered, does that mean there isn't be a response time hit compared to a typical double-buffered card?

If it's double buffered, yes.
In the case where you already get 25 ~ 30 fps, AFR is good for creating a smooth feeling. I think perhaps only the hardcore FPS gamers will be dissatisfied with 1/25 ~ 1/30 second response time. However, if the "base fps" is already in the 15 ~ 20 fps range, AFR won't help it to the "playable" status, IMHO.

 

[pcchen]

Just like ATi's old MAXX hardware used to do it, right? Voodoo did the interleaving-whatevermajiggy, right?

IIRC Voodoo's SLI makes one card renders even lines, and another card renders odd lines, then merge them to create a full frame.

NVIDIA's SLI, as explained in the slides, contains two operating modes: AFR and SFR. AFR is the MAXX thingy. SFR cut a frame in two halves for both cards to render. It's clear that AFR is simpler to operate and scales better, but when AFR failed, SFR should kick in.

Actually since sgi uses ATI's 3D chips, they also have some multi-card rendering techniques (IIRC it's quad based).

 

[Dave Baumann]

IIRC Voodoo's SLI makes one card renders even lines, and another card renders odd lines, then merge them to create a full frame.

"Digital" SLI was introduced on VSA-100; this was more like "Band Line Interleave" though since is would render alternate bands on each board, although the bands could be configurable in size.

Actually since sgi uses ATI's 3D chips, they also have some multi-card rendering techniques (IIRC it's quad based).

Eric has already pointed that out - it uses a meta / super tiling mechanism, on top of the tilining that already goes on internally anyway (and these are likely to be apportioned according to HierZ tile sizes).

 

[MfA]

The problem is that all these methods are driver hacks more than anything else, just an afterthought to push a few more boards to niche markets for which it isnt worth spending any real money on during hardware design.

 

[Kombatant]

The problem is that all these methods are driver hacks more than anything else, just an afterthought to push a few more boards to niche markets for which it isnt worth spending any real money on during hardware design.

Or, to put it with two words: Bragging rights

 

[991060]

Besides the difference on communication overhead, AFR also makes better use of vertex shader units. Basically, in AFR model, SLI works more like 2 separated cards, rather than a 6800 with 2 times the pixel shading pipelines.

 

[3dilettante]

Wasn't the previous consumer-ish level AFR solution the XGI Volari Duo?

I recall that a disconcertingly large number of situations had it performing less than the single-chip variant, and that's with the two chips on the same PCB. Perhaps it was just bad luck, if such a thing is truly what determines how an implementation fares, but at least at the consumer level things look rather mixed.

What does the driver do if a program tries to base rendering on the previous frame, give up and switch to split screen mode, or just serialize the operation?

 

[KimB]

although the response time does not improve (most people don't need much of that either).

Yes, it does. It doesn't improve as much as a "normal" increase in framerate, but response time does improve.

 

[KimB]

The problem is that all these methods are driver hacks more than anything else, just an afterthought to push a few more boards to niche markets for which it isnt worth spending any real money on during hardware design.

Right, because there isn't an SLI interconnect between the two cards, and it isn't a valid upgrade choice for those who would like to buy one today and maybe another when it's cheap later.

 

[KimB]

I recall that a disconcertingly large number of situations had it performing less than the single-chip variant, and that's with the two chips on the same PCB. Perhaps it was just bad luck

Bad luck? You see a poor implementation from XGI and you put it up to bad luck?

 

[pcchen]

Yes, it does. It doesn't improve as much as a "normal" increase in framerate, but response time does improve.

I'm curious. How does it improve response time, as the time to render a particular frame does not change?