AFR vs SLI

H

H-street

Supreme [H]ardness
Joined
Feb 25, 2005
Messages
4,538
I thought some might find this interesting..

i never really payed much attention to ATI back in the day, but i found out about ATI's attempt at "SLI" , specifically around the Geforce256/Voodoo5 days, the card was called the ATI Rage Fury Maxx and was a dual GPU solution.

Basically 2 Rage 128 chips run in the same style that the modern Nvidia/ATI setups run today, specifically using Alternate Frame Rendering (AFR)

i got my hands on one of the cards (thanks anandtech FS/T) and just thought i'd check t to see how well AFR compared to the SLI (3dfx style)..


The difference being that

Scan Line Interleve splits the image up in odd and even lines, so both cards render the same scene, just half and half.

Alternate Frame Rendering just alternates each chip on even and odd frames..

the Cards

A Pair of Voodoo2 12MBs

Voodoo5 5500 AGP

ATI Rage Fury Maxx

i decided on Quake2 as the benchmark (i need to find a good easy old DirectX benchmarking game, even old reviews used only Quake3 for benching).. This may favor the 3dfx cards due to better opengl support, but since i wanted to look at the benefit of SLI and AFR vs Single it may not matter

game was set to 800x600 / sound disabled / timedemo 1 / map demo1.dm2 (the below are reported as fps)

the Voodoo2 @ 90mhz
Single: 59
SLI : 103
75% Increase


the Voodoo2 @ 99mhz
Single: 65
SLI: 115
75% Increase

the Voodoo5 5500
Single: 125.7
SLI: 220
75% Increase

ATI Rage Fury Maxx
Single: 90.1
AFR: 129.7
43% increase

For reference a Voodoo1 scores 21fps, geforce ddr scores 178fps, 8800GT scores 522fps

i took a bunch of 8800GT single vs SLI comparisons and i came out with an average of a 45% increase (eerily similar to the 43% afr on the ati maxx).

anyone who ran an old 3dfx sli setup probably isn't suprised, the 3dfx method was rock solid.. an investment of for another 3dfx card (or in V5's case the dual GPU card) really was much more of a benefit than the AFR (even over todays SLI/Xfire setups)..


If someone has Quake2 and one of the newer Nvidia/ATI i'd love to see an 800x600 SLI vs benchmark (AFR and the Split)..

some things of note:

the ATI system has to be restarted to disable / enable AFR vs 3dfx didn't. (Similar to the early Nvidia system)
ATI's old docs suggested they had a patent on the AFR technology, wonder what ever came of that?
Alternate Frame Rendering (AFR) is a unique and proprietary technology (patent pending) for smooth 3D performance and more precise gaming control
 
From what I understand, both (current) SLI and CF support AFR. There are many modes of SLI/CF including a 50/50 split (it can also be set up to do 60/40, or 75/25, etc), AFR, checkboard, etc. I'm not sure of the exact number of possibilities, but there are quite a few.
 
4745454b said:
From what I understand, both (current) SLI and CF support AFR. There are many modes of SLI/CF including a 50/50 split (it can also be set up to do 60/40, or 75/25, etc), AFR, checkboard, etc. I'm not sure of the exact number of possibilities, but there are quite a few.
Click to expand...

but SLI (Scalable Link Interface) that we know today isn't really SLI (Scan line interleave) its more of a name recognition thing Nvidia uses to indicate dual GPU.. In fact i imagine they sat around and came up with a new meaning for SLI when they rolled it out. (and Scalable LInk interface was what they came up with)

even though SLI (3dfx) was a form of split (even vs odd) it doesn't really compare to the Split (Scissor) that is used today. Since today's SLI split is more of a top /bottom or larger chunks (which is done due to the geometry [TnL] complexity of the data being rendered)..
 
AFR is the most efficient way of using multiple GPUs. When you use scanline interleave, most triangles will have to be processed by both GPUs, causing a lot of redundant vertexprocessing and triangle setup (Voodoos still used software for this part of the rendering process, so they weren't affected by this. The driver just prepared each triangle once, then sent it to both processors). Checkerboard patterns aren't much better.
Using a 50/50 split (or any other ratio) is slightly more efficient, because triangles outside the visible rectangle can be eliminated before being drawn, but you still have a lot of redundant operations.

The main problem with AFR is that it's hard to get a smooth framerate. Often the two GPUs render the frames nearly at the same time, giving you two updates in a row, with virtually identical frames, and then you get a long wait until the next two frames. So although your framerate nearly doubles, the actual gaming experience is not much smoother.
In theory it should be possible to estimate the average frame delay based on the current framerate, and then delay the start of the next frame so that you get a smooth framerate, but as far as I know neither nVidia nor ATi are doing this...

Sadly the conclusion is that no multi-GPU rendering method is perfect. AFR has the most potential though, and I think it is also the most popular method for both nVidia and ATi at this point (for those who remember Volari, they used it too).
 
You must log in or register to reply here.
Back
Top