AGP 2x vs. 4x. Improvement or theory?

[dbal]

I saw it in another thread as a comment and I thought it's worth talking about it. Does the AGP 4x provide double the performance of a given video card that was previously working in a mobo under 2x or not? And if it doesn't why is that? Isn't 4x66 mhz a better data transfer rate than 2x66? That's what I learned at school maths longtime ago....

 

[microAmp]

I remember seeing only 1 benchmark here at AnandTech that had a Voodoo 4 or 5? You could choose a PCI version or an AGP version and Anand benched it and there was no difference.

 

[microAmp]

Linky Voodoo 5 was the card.

 

[Rand]

The V5 is not a good judge of the relative gains of PCI vs. AGP, as the V5 did not utilize any of AGP's advanced features and when utilized in an AGP slot it would effectively operate as a PCI card, with twice the available bandwidth to it.
Quite a difference from the GF2/Radeon's which took a huge performance hit in PCI.

As for the benefits of AGP4X over AGP2X.... very minimal. Your looking at roughly a 0-1% performance gain in virtually all instances, and roughly a 4-5% difference between AGP 1X and AGP 4X. Sure there is twice the bandwdith available, but considering that current graphics cards barely utilize anymore then the bandwidth that AGP 1X offers all that extra bandwidth goes to waste. And judging by the tremendous amounts of DRAM available of current consumer graphics card, bandwdith requirements from main memory to graphics cards are only going to become less and less of a factor.

Even in high end 3D modelling usage, in which DRAM requirements are stratospheric compared to most anything any of us would see, AGP 4X rarely sees beyond a 5% performance gain peaking at up to 10% over AGP 2X.

There are other benefits to AGP besides sheer bandwidth to main memory, and those benefits such as Direct mapped texture rendering, multiple address lines, higher potential power draw with much more stringent requirements etc. those benefits are what give AGP a benefit over PCI. But even AGP 1X offers most of those benefits.

But for current graphics cards the original use for AGP in that it could directly access main memory with a far greater bandwidth then PCI could offer, and that bandwidth is not shared.... is useless. With modern graphics cards offering 64-128MB in the consumer market, and 256MB + in the professional market, we simply arent touching main memory for texture storage. And if we were utilizing main memory for texture storage then even AGP 8X would be too slow to provide playable frame rates in modern games so it's pretty much a mute point.

Basically, AGP was primarily created as a work around for a problem that barely exists anymore. It's only the fringe benefits of AGP that I touched upon above that really provide any tangible boost over PCI for modern graphics cards.

 

[oldfart]

Nice post Rand. You explained it very well.

 

[jiffylube1024]

<< AGP 2x vs. 4x. Improvement or theory?
>>



AGP4X is an improvement. In theory...

 

[BFG10K]

Basically, AGP was primarily created as a work around for a problem that barely exists anymore.

Not really. Its sole purpose was to speed up the transfer of the data between the system memory to the video card and it does just that. Granted the reason for doing this has changed but it is still needed.

The next generation of games (Unreal2, Doom3 etc) which fully support vertex & pixel shaders will produce masses and masses of data that will have to be passed across AGP transfer line and AGP x4 will show a difference over AGP x2. Indeed, Intel's AGP x8 effectively doubles AGP x4's bandwidth and has been optimised for streaming transfers rather than for burst transactions for this very reason, and to also cope with exploding texture sizes. Also while you still always want to avoid transferring textures across AGP and instead keep them in the video card's local VRAM, it's still much better to have AGP x8 over AGP x1 in the occasional situation when you don't have a choice and you're forced to do it.

But right now I agree, AGP x4 shows no real difference over AGP x2 except in theoretical AGP transfer benchmark tests and very high-end 3D rendering which transfers a lot of vertex data.

 

[dbal]

<< Sure there is twice the bandwdith available, but considering that current graphics cards barely utilize anymore then the bandwidth that AGP 1X offers all that extra bandwidth goes to waste. >>



Sorry, a syntax error or sth in the sentence? Plz rewrite...



<< But for current graphics cards the original use for AGP in that it could directly access main memory with a far greater bandwidth then PCI could offer, and that bandwidth is not shared.... is useless. With modern graphics cards offering 64-128MB in the consumer market, and 256MB + in the professional market, we simply arent touching main memory for texture storage. >>



Rand, I totally respect your explanation but AGP 4x is all about data transfer rate right? Why are we talking about memory size for texture storage then?
Above all, I think it's a very nice discussion starting here...!

 

[AndyHui]

Please read the FAQ: What is the performance difference between AGP1X, 2X and 4X?

 

[Rand]

<< Sorry, a syntax error or sth in the sentence? Plz rewrite... >>



Not sure what was wrong with the sentence besides a minor typo, but I'll try to clarify the sentence....:

AGP 4X, does offer twice the bandwidth to main memory that AGP 2X is capable of, but presently available graphics cards are barely utilizing any more then what AGP 1X is already capable of. Hence all that extra bandwidth that AGP 4X is capable is wasted as it's not needed.



<< But for current graphics cards the original use for AGP in that it could directly access main memory with a far greater bandwidth then PCI could offer, and that bandwidth is not shared.... is useless. With modern graphics cards offering 64-128MB in the consumer market, and 256MB + in the professional market, we simply arent touching main memory for texture storage.

Rand, I totally respect your explanation but AGP 4x is all about data transfer rate right? Why are we talking about memory size for texture storage then?
Above all, I think it's a very nice discussion starting here...! >>



I mention texture storage because with limited texture storage on the graphics card there is an increased reliance on AGP bandwidth.... however with the available DRAM on modern cards we simply don't need to utilize the bandwidth that AGP offers.



<< Not really. Its sole purpose was to speed up the transfer of the data between the system memory to the video card and it does just that. Granted the reason for doing this has changed but it is still needed. >>



That's just the point... it's not really needed anymore. Presently available graphics cards have plenty of texture storage, and so we don't need to rely upon main memory for storage of textures. The bandwidth to main memory is there if we ever needed it... but we very seldom to need it, and whn we do it;s in limited amount and quick bursts.



<< The next generation of games (Unreal2, Doom3 etc) which fully support vertex & pixel shaders will produce masses and masses of data that will have to be passed across AGP transfer line and AGP x4 will show a difference over AGP x2. >>



They've been saying that since the original GeForce, originally it was said for T&L, then for FSAA.... now for vertex shaders. Though admittedly the argument is certainly more valid for future uses of vertex shading and some forms of AA then it ever was for initial T&L implementations.
Initial implementations of vertex shading isnt going to dramatically increase the need for texture storage, first implementations are strongly likely to utilize vertex shading to improve performance over existing cards but arent going to make significant use of new rendering capabilities on a game wide scale. Current graphics cards wouldnt be able to handle anything like that, the present capabilities limit us to vertex and pixel shading of optimizations to improve rendering of existing scenes but don't have the sheer power necessary to make game wide heavy usage of new methods of rendering that vertex shading opens up.
By the time we heavily utilize vertex shading on the scale necessary to truly dramatically increase texture storage we're looking into the future a good 3-5 years at least.
The next generation of games, arent going to nearly approach the level that would be required to put a heavy load on the AGP bus.
And we should be thankful their not... because even the bandwidth AGP 8X offers is most definitely NOT going to be enough to maintain reasonable frame rates on upcoming games if we approach any significant dependence upon it.

 

[dbal]

If I understand then, the AGP aperture size option from the bios is mostly unusable right? Because some say that it should be set to half the amount of system memory....

 

[BFG10K]

Presently available graphics cards have plenty of texture storage, and so we don't need to rely upon main memory for storage of textures.

All the texture memory in the world isn't going to do anything for the extra data that is required to be tramsmitted to the video card from the CPU when performing vertex and pixel shader operations. What you need is a high bandwidth pipe and AGP x8 will provide just that. Forget about textures because that is not the main use for AGP anymore, like I said before.

They've been saying that since the original GeForce, originally it was said for T&L

And they were 100% correct. Take a look at some benchmarks of a GF2 MX PCI or a Radeon PCI compared to their AGP counterparts in CPU limited situations and you'll see what a large difference even AGP x1 makes over PCI. Then take a look at a Kyro2 or a Voodoo5 in the same situation and you'll see an extremely minimal performance difference between the two interfaces because neither card has T&L.

Initial implementations of vertex shading isnt going to dramatically increase the need for texture storage,

No, not for the textures. It's for the extra data passed to the video card to perform those vertex/pixel operations. With programmable T&L the CPU has to send more data to the card than with fixed T&L in order for the video card to offload more calculations from the CPU. And in order to keep those rendering units feed with data a high bandwidth bus capable of high streaming performance is exactly what is needed and that is exactly what AGP x8 is for.

because even the bandwidth AGP 8X offers is most definitely NOT going to be enough to maintain reasonable frame rates on upcoming games if we approach any significant dependence upon it.

AGP x8 will do just fine for programmable T&L's bandwidth requirements just like AGP x4 does just fine for current fixed function T&L calculations.

 

[Rand]

And they were 100% correct. Take a look at some benchmarks of a GF2 MX PCI or a Radeon PCI compared to their AGP counterparts in CPU limited situations and you'll see what a large difference even AGP x1 makes over PCI. Then take a look at a Kyro2 or a Voodoo5 in the same situation and you'll see an extremely minimal performance difference between the two interfaces because neither card has T&L.

Except that the benchmarks show the Kyro takes a very significant performance hot from being run in PCI rather then AGP, a hit on level with that seen with ATi and nVidia boards.
Also, even when you disable T&L on nVidia/ATi boards the large performance difference is still there, which clearly indicates it is not related to T&L.


No, not for the textures. It's for the extra data passed to the video card to perform those vertex/pixel operations. With programmable T&L the CPU has to send more data to the card than with fixed T&L in order for the video card to offload more calculations from the CPU. And in order to keep those rendering units feed with data a high bandwidth bus capable of high streaming performance is exactly what is needed and that is exactly what AGP x8 is for.

We could take an example of 3DMark 2001 that utilizes pixel and vertex shaders in all of it's game benchmarks and is almost entirely reliant upon them in the Nature scene. It clearly utilizes pixel and vertex shaders, but we are not seeing any singificantly increased reliance upon the AGP bus despite this.

 

[LongCoolMother]

yea, anand said the 3dfx never really took advantage of the AGP benefits. so the PCI and AGP perform identical almost...i used to have a voodoo 5500, actually, i just upgraded from it to a radeon 8500 this past saturday.

 

[BFG10K]

Except that the benchmarks show the Kyro takes a very significant performance hot from being run in PCI rather then AGP, a hit on level with that seen with ATi and nVidia boards. Also, even when you disable T&L on nVidia/ATi boards the large performance difference is still there, which clearly indicates it is not related to T&L.

Do you have any links to any benchmarks that show this? I haven't seen this anywhere.

We could take an example of 3DMark 2001 that utilizes pixel and vertex shaders in all of it's game benchmarks and is almost entirely reliant upon them in the Nature scene. It clearly utilizes pixel and vertex shaders, but we are not seeing any singificantly increased reliance upon the AGP bus despite this.

(1) You don't have AGP x8 to compare to AGP x4, therefore you can't assert that AGP x4 isn't already the bottleneck.
(2) 3DMark is highly unlikely to have anywhere near the polygon counts that the likes of Doom3 and Unreal2 will have.
(3) 3DMark's results are theoretical at best and in no way translate into the real world.

 

[Rand]

Do you have any links to any benchmarks that show this? I haven't seen this anywhere.


Not off-hand but it's been chown in a few places and my own tests seem to back this up.

(1) You don't have AGP x8 to compare to AGP x4, therefore you can't assert that AGP x4 isn't already the bottleneck.
(2) 3DMark is highly unlikely to have anywhere near the polygon counts that the likes of Doom3 and Unreal2 will have.
(3) 3DMark's results are theoretical at best and in no way translate into the real world.

3DMark is intended to put more stress on the graphics engine then real world games do, and we've already seen that AGP 4X provides precious little gain over AGP 2X in 3DMark, so it's safe to say that AGP 8X isnt going to change that any.

Ah well, it's a point of debate, we shall see when the next generation of games comes along but I personally sincerely doubt we shall see AGP 8X provide much of a boost over AGP 4X in the typical uses of a desktop system.