Just HOW MUCH bandwidth do we use of the AGP interface?

[AgaBoogaBoo]

Time to time we have to remind people that right now AGP 8x is pure marketing and that it doesn't make any difference. I'm wondering if there are any numbers of as to what percent of AGP8x and AGP4x we currently use.

 

[everman]

Not sure on exact numbers, but with cards having 256megs now we hardly need much agp bandwith at all. I doubt we'll really need 8x for several years.

 

[Lord Evermore]

The only way to test it would be to use a single card in 2X, 4X and 8X modes to see whether performance changes, with very high resolution and refresh rates, high polygon counts, et cetera. I don't think you can measure bus utilization via software; memory bus bandwidth can be measured because you send in data then pull it back out, but the GPU never really sends anything back to the CPU, and isn't designed to "report" on its own performance.

Although the AGP texturing is completely unneccessary, despite that having been one of the major reasons for creating it, the bus speed is still a feature that makes a difference. With a lot of data, the bus needs to be fast enough to get the information to the GPU, even though the GPU doesn't need to make use of the bandwidth to get access to the system memory. Only with the most recent and upcoming games is 4X likely to come to anything like high utilization, due to the high polygon count and high resolutions.

Since dedicated access to the memory is primarily unneeded, and simple high-bandwidth is the major need of a video card, video will eventually move to PCI-Express or whatever successor to PCI becomes dominant. As mentioned in I think the AT 865PE article, Intel will be including a PCI-Express port for video in Grantsdale.

 

[AgaBoogaBoo]

Originally posted by: Lord Evermore
The only way to test it would be to use a single card in 2X, 4X and 8X modes to see whether performance changes, with very high resolution and refresh rates, high polygon counts, et cetera. I don't think you can measure bus utilization via software; memory bus bandwidth can be measured because you send in data then pull it back out, but the GPU never really sends anything back to the CPU, and isn't designed to "report" on its own performance.

Although the AGP texturing is completely unneccessary, despite that having been one of the major reasons for creating it, the bus speed is still a feature that makes a difference. With a lot of data, the bus needs to be fast enough to get the information to the GPU, even though the GPU doesn't need to make use of the bandwidth to get access to the system memory. Only with the most recent and upcoming games is 4X likely to come to anything like high utilization, due to the high polygon count and high resolutions.

Since dedicated access to the memory is primarily unneeded, and simple high-bandwidth is the major need of a video card, video will eventually move to PCI-Express or whatever successor to PCI becomes dominant. As mentioned in I think the AT 865PE article, Intel will be including a PCI-Express port for video in Grantsdale.

I think I'll try that once I get my new system which I'll order in a week or two because it won't make a difference for my Radeon VE!

 

[InfraggableKrunk]

THERES A MARGINAL DIFFERENCE , do you want to know how marginal, most people use a ti4800 over a ti4600 in bencmarking and leave out the ti4600, why Because its redundant. Sure there is a difference but not much

I say were only usin still only about 3.5 agp instead of agp4x, ( ok abstract but thats how i look at it) with a 8x card and 8x mobo youll probably be getting a 3.55 agp, so its really pointless

until someone learns how to use agp 8x im gonna stick with my 4x motherboard

 

[AgaBoogaBoo]

Originally posted by: InfraggableKrunk
THERES A MARGINAL DIFFERENCE , do you want to know how marginal, most people use a ti4800 over a ti4600 in bencmarking and leave out the ti4600, why Because its redundant. Sure there is a difference but not much

I say were only usin still only about 3.5 agp instead of agp4x, ( ok abstract but thats how i look at it) with a 8x card and 8x mobo youll probably be getting a 3.55 agp, so its really pointless

until someone learns how to use agp 8x im gonna stick with my 4x motherboard

What are you talking about? Getting a 3.55 instead of a 3.5? That doesn't matter, if it can go 4x, it'll use as much bandwidth as possible. Also, if both of them support more than what is needed, what makes one faster than the other?

 

[AgaBoogaBoo]

Double Post, delete please!

 

[InfraggableKrunk]

it'll use as much bandwidth as possible

HAH

you know most pc parts cant use all their allocated bandwith

cough, cough usb 2, cough, cough

i mean to think that all producst " use as much bandwidth as possible" is a farse

hmm time to research this

the 3.5 thing was an abstract way of showing the true usage of a 4x and the true usage of an 8x

 

[RaynorWolfcastle]

Originally posted by: InfraggableKrunk
it'll use as much bandwidth as possible
HAH
you know most pc parts cant use all their allocated bandwith
cough, cough usb 2, cough, cough
i mean to think that all producst " use as much bandwidth as possible" is a farse
hmm time to research this
the 3.5 thing was an abstract way of showing the true usage of a 4x and the true usage of an 8x

You're confusing a few different concepts, let me clear this up for you:

- There is always overhead for any given interface. That is, even though your bus can handle 12 Mbit/s (USB 1.1), there's a portion of that bandwidth that is used for various system tasks such as ensuring that the device is still connected, parity bits (for error checking), etc. So though you may have 12 Mbit/s bandwidth you only have say 10 MBit/s of usable bandwidth to you.

- Any controller, no matter the interface (USB2, Firewire, IDE/ATA, chipset etc.) does not work with 100% efficiency. That is you will never use up that 10 Mbit/s available to you because of things like the controller's internal latencies. This depends on the interface's specifications and on the chip's design as both contribute to how easy or difficult it is to minimize these inefficiences. Case in point is USB2 vs Firewire, FireWire controllers and devices are more efficient than their USB2 counterparts so it is faster despite having a lower theoretical throughput.

- If an application requires a certain amount of throughput, it does not matter how the controller provides it as long as it does, ie whether you plug your mouse on a USB2 port or a USB 1.1 port won't much matter since the mouse won't exceed 12 Mbit/s anyway.


What does this mean for AGP, you ask?

Well for one thing the increasing amount of onboard texture memory on today's video cards mean that nearly no textures are repeatedly being passed over the AGP port anyway, and textures are what is most bandwidth demanding. Instead, most of what is being passed over is "information" that the card needs to render the scene. This may be a little confusing because I am trying to simplify this (maybe a little too much). Anyhow, this rendering information takes a lot less bandwidth than constant swapping of huge textures ofver the AGP port.

So let's say AGP 4x theoretically provides 1066 MB/s while AGP 8x provides 2133 MB/s, there will be no difference between the two if the application only requires 500 MB/s. Even taking overhead and inefficiencies into account AGP 4x would still be able to provide ample bandwidth for that application, so you would see very little difference between an AGP 8x controller and an AGP 4x one. The only reason AGP 8x may be marginally faster is because newer AGP controllers may have reduced latencies compared to older AGP 4x controllers.

I hope that clears things up

 

[ViRGE]

Originally posted by: Lord Evermore
Since dedicated access to the memory is primarily unneeded, and simple high-bandwidth is the major need of a video card, video will eventually move to PCI-Express or whatever successor to PCI becomes dominant. As mentioned in I think the AT 865PE article, Intel will be including a PCI-Express port for video in Grantsdale.

Personally, I find it wierd that they're going back to a shared bus with video cards. Besides AGP texturing and DMA, one of the benefits of getting video cards off the PCI bus was that the video card was no longer sucking up the bus's bandwidth, freeing it up for sound cards, HD controllers, NIC's, etc. It seems silly to go back to a shared bus when it has gotten to the point where Intel has created CSA and other non-PCI links for NICs and HD's in order to keep from saturating the bus anyhow.

What is the PCI Express bus rated at? Unless it's significantly faster(on the order of several gigabits a second), I can't see this being a good thing.

 

[Tab]

16-Bit PCI-X which will replace AGP8x has a bandwidth of 4GB/s and feeds your video card 60-Watts, which is still not enough for your 86-Watt 9700 Pro.

 

[zephyrprime]

I've always wondered about this so I thought I would put it to the test.

I've got a radeon 9500 hacked into a 9700 and I'm running with a athlon XP at 1875mhz.
I tried 4x agp and 2x agp. I haven't tried 1x. Other sites have already tested the difference between 8x and 4x and it's ridiculously small.

Anyway, I ran 3dmark 2001 and the results are:
4x: 12217
2x: 12204

So small that it's probably within the margin of error.

Now, I also took a look at what 3dmark had to say about the high polygon count test. 3dmark said it was pushing 59.1 MT/sec during the high poly test with 1 light.
Assuming that the mesh averages out to a hexagonal one, that would mean 1 vertext connects with 6 triangles and since each triangle has 3 vertices, each triangle averages 1/2 of a vertices. Assuming that each vertex takes 4x32bit = 16bytes, this would work out to ~472MB/sec which is only a little less bandwidth than agp 2x can provide.

Now I made a lot of assumptions about things I don't really know a whole lot about so take this all with a cow lick of salt.

 

[ViRGE]

So that's 4 gigabytes, right? As for the power, routing that kind of power is approaching silly levels. It's bad enough with the CPU, no need to do it for cards too when we can seem to get away with direct connections to the PS.

 

[zephyrprime]

which is still not enough for your 86-Watt 9700 Pro.

Are you sure about that? That's more power than any CPU and the 9700pro has a really small HSF.

 

[BmXStuD]

dunno my new asus will have 8x agp even know my gf4 mx is 8x i doubt it will even mater.

 

[AgaBoogaBoo]

Originally posted by: zephyrprime
I've always wondered about this so I thought I would put it to the test.

I've got a radeon 9500 hacked into a 9700 and I'm running with a athlon XP at 1875mhz.
I tried 4x agp and 2x agp. I haven't tried 1x. Other sites have already tested the difference between 8x ans 2x and it's ridiculously small.

Anyway, I ran 3dmark 2001 and the results are:
4x: 12217
2x: 12204

So small that it's probably within the margin of error.

Now, I also took a look at what 3dmark had to say about the high polygon count test. 3dmark said it was pushing 59.1 MT/sec during the high poly test with 1 light.
Assuming that the mesh averages out to a hexagonal one, that would mean 1 vertext connects with 6 triangles and since each triangle has 3 vertices, each triangle averages 1/2 of a vertices. Assuming that each vertex takes 4x32bit = 16bytes, this would work out to ~472MB/sec which is only a little less bandwidth than agp 2x can provide.

Now I made a lot of assumptions about things I don't really know a whole lot about so take this all with a cow lick of salt.

Wow, looks like even if so meone runs at 2x, they may not notice a difference. If you don't mind, try it at 1x and 8x as well. I would but its not really going to make a difference on my Radeon VE SDRAM, but I'm getting a 9700 Pro in 2 weeks or so

 

[squidman]

Nah, there aint no numbers for that. Just sole opinions. Truth is, we really dont use agp 8x...yet. That explains why GF4600 is hella fast (its AGP4x) and is quite comparable to Radeon 9700...IF GF4600 had 256 memory interface, it would be HELLA fast. Just like the others, ill throw my assumption of AGP bandwidth usage: 700mb/s to 1000mb/s. Thas it.

 

[InfraggableKrunk]

Originally posted by: zephyrprime
I've always wondered about this so I thought I would put it to the test.

I've got a radeon 9500 hacked into a 9700 and I'm running with a athlon XP at 1875mhz.
I tried 4x agp and 2x agp. I haven't tried 1x. Other sites have already tested the difference between 8x ans 2x and it's ridiculously small.

Anyway, I ran 3dmark 2001 and the results are:
4x: 12217
2x: 12204

So small that it's probably within the margin of error.

Now, I also took a look at what 3dmark had to say about the high polygon count test. 3dmark said it was pushing 59.1 MT/sec during the high poly test with 1 light.
Assuming that the mesh averages out to a hexagonal one, that would mean 1 vertext connects with 6 triangles and since each triangle has 3 vertices, each triangle averages 1/2 of a vertices. Assuming that each vertex takes 4x32bit = 16bytes, this would work out to ~472MB/sec which is only a little less bandwidth than agp 2x can provide.

Now I made a lot of assumptions about things I don't really know a whole lot about so take this all with a cow lick of salt.

THANK YOU

thank you for confirming my findings , i knew that agp 4x wasnt even been used to its max

although i never knew it was as low as 2x ! damn there a lot of headroom even with 4x agp

 

[m1ke101]

someone should try 1x I'll try when I get back home

 

[Mavrick007]

When AGP 4x first came out, there was alot of people who were questioning the necessity of it then since alot had AGP 2x boards.
It was found that with the then current apps and games, 2x was enough and 4x wasn't going to be saturated for a long while to come.

 

[zephyrprime]

Alright, tried 1x. Got 12122 on 3dmark2001. So the results are:

4x: 12217
2x: 12204
1x: 12122

In the high poly test with 1 light, the results were:
4x: 59.1 Megatriangles
2x: 57.8
1x: 57.8

I can't test 8x since my MB only supports 4x.

So, basically, we need at least AGP 1x for the high polygon games of tomorrow. Better run out and upgrade right now!

ps. I previously wrote "... difference between 8x ans 2x ..." but that was a typo. I really meant "... difference between 8x and 4x ..."

 

[FishTankX]

Originally posted by: ViRGE

Originally posted by: Lord Evermore
Since dedicated access to the memory is primarily unneeded, and simple high-bandwidth is the major need of a video card, video will eventually move to PCI-Express or whatever successor to PCI becomes dominant. As mentioned in I think the AT 865PE article, Intel will be including a PCI-Express port for video in Grantsdale.

Personally, I find it wierd that they're going back to a shared bus with video cards. Besides AGP texturing and DMA, one of the benefits of getting video cards off the PCI bus was that the video card was no longer sucking up the bus's bandwidth, freeing it up for sound cards, HD controllers, NIC's, etc. It seems silly to go back to a shared bus when it has gotten to the point where Intel has created CSA and other non-PCI links for NICs and HD's in order to keep from saturating the bus anyhow.

What is the PCI Express bus rated at? Unless it's significantly faster(on the order of several gigabits a second), I can't see this being a good thing.

PCI express is like Serial ATA,in which each card get's dedicated bandwidth from the controller. I believe the intitial figures are something like 266MB/s for each port.

 

[FishTankX]

They're talking about linking like, 8 ports together for graphics cards or something, though. So a single card can use multiple ports, if I understand this correctly.

 

[Viper96720]

So voodoo like sli setups maybe possible with pci express? If the video card companies makes a card that can do it.

 

[RaiderJ]

It sounds like higher-end graphics cards don't need the AGP bandwidth for the most part, but do onboard graphics see a larger benefit from the faster AGP standards? Anyone have onboard graphics and care to test this out?

Do onboard graphics even have their own RAM, or is it always shared with system memory?