AGP 4x vs AGP 8x

[Lyquist]

I just ran a few benchmarks on my machine at 8x and 4x and saw no difference...even in Doom3. It would seem that nothing takes advantage of 8x yet. I ended up doing the benchmarks because I was having stability issues with my card at 8x...changing to 4x in the ATI drivers seems to have solved the problems that I was having.

Lyquist

 

[Gillette]

No card even uses close to the bandwidth offered by AGP 4x... hence my inability to put up with PCI-E trolls.

 

[pxc]

No card even uses close to the bandwidth offered by AGP 4x...

You can use synthetic benchmarks to show the increase in performance, so it's not correct to say that. It's not even correct to say that no software uses the bandwidth offered by AGP4x. A more narrow case, that no shipping games use use the bandwidth offered by AGP4x, would be more correct. But why would a game maker require AGP8x? That's just as bad as requiring a single GPU. It's a design decision, not a limitation of the cards.

LOL @ your signature. Your AGP X800XT is @ a X850XT PE (native PCI-E). Nice trick. How did you get the 4GB/s GPU->bus bandwidth and extra 2GB/s bus->GPU bandwidth?

 

[Gillette]

You can use synthetic benchmarks to show the increase in performance, so it's not correct to say that. It's not even correct to say that no software uses the bandwidth offered by AGP4x. A more narrow case, that no shipping games use use the bandwidth offered by AGP4x, would be more correct. But why would a game maker require AGP8x? That's just as bad as requiring a single GPU. It's a design decision, not a limitation of the cards.

LOL @ your signature. Your AGP X800XT is @ a X850XT PE (native PCI-E). Nice trick. How did you get the 4GB/s GPU->bus bandwidth and extra 2GB/s bus->GPU bandwidth?

hence my inability to put up with PCI-E trolls.

On another note, it's refering to clock speed nublar.

AGP 4x is in no way a bottleneck for any currently existing GPU.

"That's just as bad as requiring a single GPU" as opposed to what? Not requiring a GPU at all

"You can use synthetic benchmarks to show the increase in performance, so it's not correct to say that." Why on earth would you use a synthetic benchmark over a real one? Complete waste of time if you ask me

"But why would a game maker require AGP8x?" Really? This is news to me! I certainly can't say I've ever seen "AGP 8X" as a system requirment on a game.

"It's not even correct to say that no software uses the bandwidth offered by AGP4x. A more narrow case, that no shipping games use use the bandwidth offered by AGP4x, would be more correct." Again, news to me... Only bottlenecking I'm seeing is caused by the GPU its self, not the bus.

And, on a final note, I <3 how my points are being flamed by someone who owns 2 Dells

 

[pxc]

On another note, it's refering to clock speed nublar.

AGP 4x is in no way a bottleneck for any currently existing GPU. Also... "That's just as bad as requiring a single GPU" as opposed to what? Not requiring a GPU at all

Nice to know what you actually meant since you're equating an AGP card with a native PCI-E card in your signature.

ATI and Intel seem to believe that AGP4x is a bottleneck. As a matter of fact, ATI trashed NVIDIA for not having a native PCI-E interface because they said a bridge (nvidia's HSI) wasn't fast enough to do real-time HDTV encoding (streams the data back to memory from the GPU). Intel shows performance of multiple 720p stream decoding running 4 times faster on PCI-E (no CPU decoding) than AGP8x (and 5x faster than AGP4x) because of the higher GPU->host bandwidth available.

In case you misunderstood what I was saying above, game makers aren't going to require/rely on AGP8x bandwidth when they release a game for the same reason that they don't write for a single GPU: it limits the target audience. Due to the high total bandwidth available on PCI-E (4GB each way at the same time), you will see some newer types of effects and applications in the future that take advantage of it. See ATI's PCI-E page for an example.

edit since you added this:

And, on a final note, I <3 how my points are being flamed by someone who owns 2 Dells

And...? I flamed your signature, not your points. You made a sweeping statement that is easily proven false. You took one case (game benchmarks) and applied that to the bus/card.

 

[Gillette]

Intel shows performance of multiple 720p stream decoding running 4 times faster on PCI-E (no CPU decoding) than AGP8x (and 5x faster than AGP4x) because of the higher GPU->host bandwidth available.

Using what GPU?

...for the same reason that they don't write for a single GPU: it limits the target audience.

Again, show me a game that is designed for dual GPU use

Due to the high total bandwidth available on PCI-E (4GB each way at the same time), you will see some newer types of effects and applications in the future that take advantage of it.

Yes, once cards start using enough bandwidth to make it worth while.

 

[pxc]

Again, show me a game that is designed for dual GPU use

I'll hold back saying what I want to say.

Writing for a single GPU as in for ONLY a 6800 (requires SM3.0 or NVIDIA only OpenGL extensions), ONLY a x800 (requires 3Dc or ATI only OpenGL extensions) or ONLY a Voodoo5 (GLide w/VSA100-only features).

And yeah, I know NVIDIA added 3Dc support in more recent drivers and ATI will support SM3.0 in the R520.

 

[Gillette]

Thanks for rephrasing that.

Here's another brainbuster for you. If synthetic benchmarks are such a great comparison and show the difference between AGP and PCI-E performance why does my system score the same as my friend's 3.4EE X800XT PE PCI-E system in 3DMark03?

 

[Lyquist]

I guess my point is that I don't see a difference in any applications that I actually use. I don't do any HDTV decoding or encoding, so I don't worry about that. I basically use my computer for my business and the occasional game of Doom3 or Starcraft.

Lyquist

 

[Gillette]

I should hope you don't do any HDTV decoding, afaik it's still illegal in most countries

 

[pxc]

Here's another brainbuster for you. If synthetic benchmarks are such a great comparison and show the difference between AGP and PCI-E performance why does my system score the same as my friend's 3.4EE X800XT PE PCI-E system in 3DMark03?

It depends on what the synthetic benchmark is measuring, of course. 3DMark03/05 measure 2 things: shader performance and CPU rendering performance. Both have little to do with bus bandwidth. If you were to put your own system in AGP0x mode (PCI-66), AGP1x, AGP2x, AGP4x and AGP8x modes and run the benchmarks, you would see a slight difference in scores.

But if you were to run a benchmark that is bottlenecked by AGP speed (3DMark2001, for example), you would see a bigger difference on your own system between AGP0x-AGP8x. Comparing scores between different systems, especially with different CPUs and much different chipsets doesn't really say much.

 

[Scali]

Thanks for rephrasing that.

Here's another brainbuster for you. If synthetic benchmarks are such a great comparison and show the difference between AGP and PCI-E performance why does my system score the same as my friend's 3.4EE X800XT PE PCI-E system in 3DMark03?

Simple, 3DMark03 isn't limited by the bandwidth of the AGP bus. Why not? It does nearly everything on the GPU, and the bus is mainly used for sending a handful of draw-calls every frame.
You will only get an advantage from a faster bus if the slower bus was not fast enough to begin with.

 

[Gillette]

But if you were to run a benchmark that is bottlenecked by AGP speed (3DMark2001, for example), you would see a bigger difference on your own system between AGP0x-AGP8x.

The only limiting factor in 3DMark2001 at this point is CPU, that's common knowledge.

Comparing scores between different systems, especially with different CPUs and much different chipsets doesn't really say much.

So then how the hell does your argument have any substance

You will only get an advantage from a faster bus if the slower bus was not fast enough to begin with.

Exactly, and AGP isn't too slow.

 

[Scali]

Exactly, and AGP isn't too slow.

Correction: it is not too slow for most games. It is certainly too slow for high-quality video encoding, especially for multiple streams.
It is also not a full-duplex solution, so it is too slow for any GPU->CPU transfer of data in general. PCI-e opens up a whole new range of applications for GPUs this way.

 

[Gillette]

Yes, there is a whole new realm of expansion available, but it is relatively meaningless until hardware is made that can actually take advantage of it.

 

[Scali]

Yes, there is a whole new realm of expansion available, but it is relatively meaningless until hardware is made that can actually take advantage of it.

It's not the hardware that has to take advantage, it's the software!
Software has always avoided reading from videomemory with the CPU because it was horribly slow on the PCI/AGP buses.

 

[ThomasE66]

Yes, there is a whole new realm of expansion available, but it is relatively meaningless until hardware is made that can actually take advantage of it.

Not just GPU hardware either. Today's highest end GPU solutions are basically becoming CPU bound in what they can accomplish on all but the fastest of today's CPU's. If the GPU/Graphics driver is waiting on the CPU, it's hardly going to be able to take advantage of enormous bandwidth. The foundation is there though, and as video coprocessors of various types become more commonplace we may just see a sudden emphasis on bus speeds as a bottleneck.

Then there is software. If an app isn't written to take advantage of loads of System->GPU bandwidth, then it's being wasted. Again, it's a foundation for future development.

It's funny; I recall this same arguement only it was PCI vs AGP 2x These changes, including PCI-Ex, are evolutionary rather than revolutionary. We just move a potential future bottleneck further down the list with PCI-Ex, as well as open doors to new applications.

I doubt you'll ever see an application that shows a huge difference between AGP 8x and AGP 4x. It's possible that it will happen down the road when games require ever increasing texture sizes, but by then you won't be getting high end GPU's on AGP of any flavor most likely.

 

[Gillette]

I <3 you.

 

[slowbiz]

I guess my point is that I don't see a difference in any applications that I actually use. I don't do any HDTV decoding or encoding, so I don't worry about that. I basically use my computer for my business and the occasional game of Doom3 or Starcraft.

Lyquist

Agreed. Who needs their computer to split the atom? I would rather spend thousands of dollars on musics.

 

[rcolbert]

New technology needs a critical hardware mass before it's fully leveraged. Can you imagine Microsoft releasing 64-bit x86 Windows without any processors in the market that could actually use it? I think the consensus is that if you're building a new PC that you should consider PCI-E as the forward-looking thing to do. Whipping out benchmarks and use-cases is missing the point.

 

[Gillette]

I've got nothing wrong with PCI-E, I guaruntee you that my next PC (whenever I get one) will be a PCI-E system.

By all respects AGP is dead now, I won't argue that.

 

[ThomasE66]

AGP is by no means dead. I just has a terminal illness at this point

 

[Goat]

AGP is by no means dead. I just has a terminal illness at this point

ya.....marketing. lol. shinny new PCI-E is easier to sell for more money.

 

[rcolbert]

ya.....marketing. lol. shinny new PCI-E is easier to sell for more money.

Goat was scoffing when PCI was first introduced, claiming that ISA was fast enough, EISA if you wanted to build a server, but certainly Microchannel was the real deal once IBM decides to ease up on the licensing costs. They're always trying to push that new technology on us that we don't really need. I bet HL-2 would run just fine on a 486 except that Valve intentionally cut it's balls off my making it only work on those shiny new Pentium class processors.

 

[ThomasE66]

Goat was scoffing when PCI was first introduced, claiming that ISA was fast enough, EISA if you wanted to build a server, but certainly Microchannel was the real deal once IBM decides to ease up on the licensing costs. They're always trying to push that new technology on us that we don't really need. I bet HL-2 would run just fine on a 486 except that Valve intentionally cut it's balls off my making it only work on those shiny new Pentium class processors.

I hope you meant 8 bit ISA. Everyone knows that 16 bit ISA slots are useless; the extra 8 bits are just gratuitous marketing.

 

[sculelos]

Actually I believe its easier to make PCI-Express cards more effieciently if thats all you have to make so when possible everyone should make the switch, but then I will never buy another AGP system, only PCI, but I am using AGP right now, and RIGHT now thier is literally no difference in AGP Verses PCI-Express(Except for movies maby), So just get PCI-E next time you buy a system because they are also cheaper, so just get one.