AOA Forums AOA Forums AOA Forums Folding For Team 45 AOA Files Home Front Page Become an AOA Subscriber! UserCP Calendar Memberlist FAQ Search Forum Home

Go Back   AOA Forums > Hardware > Graphics and Sound cards; Speakers and other Peripherals

LinkBack Thread Tools Rate Thread
  #81 (permalink)  
Old 3rd March, 2005, 01:36 AM
Member/Contributor/Resident Crystal Ball
Join Date: March 2004
Posts: 7,451

So, here ya go.

For AFR, the driver renders all of frame n on gpu 0, and all of frame n=1 on gpu 1. Frame n=2 renders on gpu and son. As long as each frame is self-contained(that is, frames share little or no data) AFR is maximally efficient, since all the rendering work, such as per vertex, rasterization, and per pixel work split evenly across GPU's. IF some data is shared between the frames(for example, re-using previous rendered-to textures), the data needs to transfer between the GPUs. This data transfer constitues comminunications overhead preventing a full 2x speed-up.
This is similar to 3dfx's version, although entirely different. instead of lines, it's full frames.

For SFR the driver assigns the top portion of a frame to GPU 0 and the bottom portion to gpu 1. The size of the top versus the bottom portion of the frame is load balanced: if GPU 0 is underutilized in a frame, becasue the top portionis less work to render than the bottom portion, the driver makes the top portion larger in an attempt to keep both GPUs equally busy. Clipping the scene to the top and bottom portions for, respectively, GPU 0 and 1, attempt to avoid processing all vertices in a frame on both GPUs.

SFR mode still requires data sharing, for example, for render-to-texture operations. Becasue AFR generally has less communications overhead and better vertex-load balancing than SFR, AFR is the preffered mode. Sometimes, however, AFR fails to apply, for example, in an application limits the maximum number of frames buffered to less than 2
So there you have's not hardware, except for the wee bit of vertex and render to texture data sharing. This is the only time that the pcb interconnect is used, besides passing the rendered frames to the on-chip memory on the primary gpu, and then to the framebuffer on the primary card. Everything that happens in SLi is dependant on the driver, and how it sends the vertex data streams from DirectX to each card, whether it be half the frame, or every second frame of data. The Nvidia chipset is responsible for duplication of the data down both "converted" 8x links, or the different data, depending on what the driver says. Regardless of how you look at it, whether it be the wrong or right way, SLi is still driver dependant, even from nVidia's own mouth.
Digg this Post!Add Post to del.icio.usBookmark Post in TechnoratiFurl this Post!
Reply With Quote
  #82 (permalink)  
Old 3rd March, 2005, 10:44 AM
Chief Systems Administrator
Join Date: September 2001
Location: Europe
Posts: 13,075

Originally Posted by cadaveca forget i have the gpu programming guide infront of me? Not explaining it for me, but for the benefits of others. Devil's advocate. Suits me well, sometimes.
The GPU programming guide only provides software implementation details. I've found it to be comprehensively incomplete in description of SLi; it doesn't touch on the hardware side. However, there are some other presentations that do detail the hardware, but I don't have them on my work machine.
Any views, thoughts and opinions are entirely my own. They don't necessarily represent those of my employer (BlackBerry).
Digg this Post!Add Post to del.icio.usBookmark Post in TechnoratiFurl this Post!
Reply With Quote
  #83 (permalink)  
Old 3rd March, 2005, 11:03 AM
gsi31099's Avatar
Join Date: November 2004
Location: South Africa
Posts: 37


All this to prove that SLi doesn't really work well? It doesn't mean much to the average Gamer/PC Enthusiast! People coming here to find out if SLi is worth it, won't understand heads or tails from this post! At least put all this info into something worthwhile.
Koolance PC2-650-BU for CPU
Antec True550 P/Supply
AMD64 3200+ @ 2280mHz(228x10)
1.0G Ram(Kingston DDR400)
36GB WD Raptor 10,000rpm
ASUS 6600GT's SLi'ed (588/1098)
Digg this Post!Add Post to del.icio.usBookmark Post in TechnoratiFurl this Post!
Reply With Quote
  #84 (permalink)  
Old 3rd March, 2005, 11:14 AM
5 Card Draw Champion, Tomb Digger Champion, Alpha Force Champion, Autobahn Champion, Big Bird Hunting Champion, Mumu Champion, Attack of the Tweety Zombies Champion
Join Date: January 2005
Location: DORSET, UK
Posts: 2,170

No thats not true...i have read all the above and i am now perfectly happy with the fact that my understanding of these things is nil

Having said that in the words of a a certain Homer J Simpson "everytime i learn something new, something else gets pushed out" so it doesnt matter
AM2 X2 4000

Join our F@H team today..
Are you a subscriber? No? well then, why not....
Subscribe to AOA for a better experience and faster downloads!
How to unlock and overclock your Nvidia 6800
BIOS Volt-mod your 7800gt

AOA Team fah
Digg this Post!Add Post to del.icio.usBookmark Post in TechnoratiFurl this Post!
Reply With Quote

Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)
Thread Tools
Rate This Thread
Rate This Thread:

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is Off
HTML code is Off
Trackbacks are On
Pingbacks are On
Refbacks are On

All times are GMT +1. The time now is 03:10 PM.

Copyright ©2001 - 2010, AOA Forums
Don't Click Here Don't Click Here Either

Search Engine Friendly URLs by vBSEO 3.3.0