Parhelia!!!

[superfly]

Originally posted by Rags

Such an important feature as AGP 8X? That's important? Come on, get real.

Rags

You bet it is if we ever want to see scenes with huge amounts of polys onscreen and there's no point in cards makers advertising that any given card can process 130 million polys(like nvidia claims for the GF4 TI 4600),if you simply can't upload 1/10th of that to the video card,due to bandwith issues(system memory,agp bandwith,etc...),in a game environment in realtime....


Everybody says that AGP 4x is enough for now,but that's mostly due to the fact that game developers try to use effects that are either cpu bound or video card dependant,while trying to minimize as much as possible to amount of data that has to be uploaded to the video card itself...


Bus speeds are actually the main reason why we see consoles displaying huge amounts of polys on screen,much higher that what we currently have in pc games,and it isn't because if the cpu's they use or their graphics chips are more powerfull that in high end pc's,because they aren't...


And as everybody can see the attention in pc's lately has shifted from just releasing faster cpu's to increasing bus and memory speeds,at least at a much faster rate than we've seen over the last few years....

There's even initiatives to implement even faster versions of AGP(called serial AGP btw),to increase the speed to levels that go beyond AGP 8x.

[Rags]

Obviously you have no clue on what AGP8X entails.

Rags

[superfly]

Originally posted by Rags
Obviously you have no clue on what AGP8X entails.

Rags

Really?...then please...enlighten me....

[Rags]

Originally posted by superfly



Really?...then please...enlighten me....

Explain how you think AGP 8X will speed up the graphics performance. And I will show you where you are going wrong.

Rags

[superfly]

Well let's see...one of the main bandwith eaters would be transfering large amounts of vertex data through the agp bus for one,in order to make full use of the transformation engine built into cards like the GF3/GF4/Radion 8500...


I used to remember exactly how much bandwith is used up for a given amount of polys wich even took in consideration that you can optimize it in such a way as to not needing to calculate all the vertices,but i do remember that it doesn't take a huge amount,relative to what current cards can handle,to saturate the AGP bus...


There is a vertex compression scheme built into DX8 as well,and recent cards do have the ability to store some vextex data locally,within the video card memory,to avoid using the agp bus as much as possible,but that will only take you so far,since there's only so much memory available...

[Wombat]

The spec hasn't been finalized, but the industry works off of drafts to get 80 or 90% of the work done. If SiS was trying for 8x, then they'd say it, as they have. They're running a risk, that's all.

And no, AGP 8x doesn't provide needed bandwidth. People aren't even using AGP2x to the fullest, and just about anything will do that these days.

You bet it is if we ever want to see scenes with huge amounts of polys onscreen and there's no point in cards makers advertising that any given card can process 130 million polys(like nvidia claims for the GF4 TI 4600),if you simply can't upload 1/10th of that to the video card,due to bandwith issues(system memory,agp bandwith,etc...),in a game environment in realtime....

No, you only need the extra bandwidth for textures, the raw data has plenty of pipe. Still, the textures aren't using a fraction of what's available.

[superfly]

Wombat...The original function of AGP was indeed to use system memory to store textures since back then,having video cards with large amounts of memory onboard simply wasn't feasable at prices that people could afford....


That's isn't the case today,and several devlopers have commented that you only use AGP to transfer textures on the fly at a last resort and it's quite simple to prove really....Try to run a texture intensive game like max payne on the highest texture quality settings,which remedy strongly recomends a 64 meg card,on a 32 one and you'll see that when the AGP bus is used to transfer textures on the fly(no preloading),performance will take a huge nose dive...


And the vextex still problem still applies,otherwise why would microsoft develop vextex compression routines and video card makers desing their chips to store and retrieve vextex data from video card memory,if AGP 2x/4x is more that enough bandwith???


And why will upcoming games like DOOM3 and Unreal 2 use about 100.000 to 150.000 polys per frame and if we want 60 fps as an average frame rate,that comes out to 6 to 9 million polys/sec,a joke when we have video cards like the GF4 which claim to handle 130 million+...

[superfly]

Wow...this thread has become very quiet all of a sudden....


Btw,i found an example in which a game engine stores vertex data in local video card memory,to avoid overloading the AGP bus...


Anyone remember a tech demo using a game engine made by a company called Crytek,in which you controled a free roaming camera which showed a large outdoor level filled with huge amount of trees and the odd dinosaur...it was one of the demos that Nvidia hyped as using the GF3's full capabilities...


According to the developers,one of the ways scenes with that many trees was possible was to store geometry data for a couple of those trees in the video cards memory,but use the programable vertex engine in the GF3 to change their shape and size,so that you'd get alot more variety in the overall scene,without needing to constantly upload new vertex data...


Here a pic...https://www.crytek.com/hp/main_files...ery/xisle7.jpg

[Ali]

Wow, thats pretty! Wouldnt mind a game that looks like that. you wouldnt be too worried about playing the game, just wonder around looking at the view!

Ali

[superfly]

Actually there's a demo available if you want to try it,it's about 40 megs...of course anyone with a GF3 or Radeon 8500 will get the best results and even then,it'll get choppy in places if you go above 1024*768*32bit and 24 bit z-buffer...

https://www.demonews.com/files/view.php?id=101

Try it....

[Ali]

Downloading it now, wonder how it looks on a G400? Might have to try it onthe Radeon instead.

Ali

[superfly]

One thing's for sure,it would make a kickass counterstrike map,since it would nearly be impossible to spot snipers...lol.

[Pace]

Hope these are right:
AGP 2x - 266MB/s
AGP 4x - 533MB/s
AGP 8x - 1066MB/s

Anyway, I think PC133 RAM maxes out at around 1GB/s, so DDR 133 will hit 2GB/s. I suppose this makes AGP 8x more appealing, especially as the 10GB/s+ of GPUs now makes 500MB/s seem slow! But try switching between 4x and 2x if you can, see if you can notice any difference. If there isn't any gain going up to 533MB/s, then a further 533MB/s is even less likely to help. Unless SiS are planning to use AGP texturing a lot? Dunno mate

P.

[Ali]

That demo doesnt work on either the G400 or a Radeon (1st gen)

Its got a dropdown box to pick between Geforce3, or Geforce1&2

If you pick Geforce3, it says you dont have the hardware, and tels you to use Gefore1 or 2.
If you pick Geforce1&2, it just closes when you hit go (well, Fire in this case).

Pity.

Ali (sorry for being off-topic)

[Rags]

Originally posted by superfly
[B]Well let's see...one of the main bandwith eaters would be transfering large amounts of vertex data through the agp bus for one,in order to make full use of the transformation engine built into cards like the GF3/GF4/Radion 8500...

It's not right now, but it may be in the future.

I used to remember exactly how much bandwith is used up for a given amount of polys wich even took in consideration that you can optimize it in such a way as to not needing to calculate all the vertices,but i do remember that it doesn't take a huge amount,relative to what current cards can handle,to saturate the AGP bus...

It wouldn't take much to saturate the AGP bus at all. The problem with DX8 is that it is still dependent on the speed of the memory interface. Memory speed on the motherboard is a huge drawback and unless you are hip to faster Rambus, then AGP 8X is nothing but a pipedream as far as benefits go.

There is a vertex compression scheme built into DX8 as well,and recent cards do have the ability to store some vextex data locally,within the video card memory,to avoid using the agp bus as much as possible,but that will only take you so far,since there's only so much memory available...

Yes, only so much memory available is correct. But the action of the software making a request for the data, sending it down to the motherboard's memory, storing it, then the video card using AGP 8X as a means to request the stored data, fetch the data, then send the information back that the operation has completed is taking a whole bunch of memory bandwidth in order for that to even be remotely useful when compared to the software sending it to the card, the card storing and manipulating the data on board. With cards coming out with 256 MB of memory on board, this shouldn't be much of an issue and those that will perform this on board with local memory will be the ones with useful performance. There is no way, with current memory tech or even AGP 8X, the performance level will be acceptable when using the local memory to store the info.(EDIT: unless it is using execution at the motherboard's memory level, and we all know how tricky it is to perform DME with certain motherboard chipsets )


Rags