PCI Express performance versus AGP 8x in 3d gaming ?

nevyn

Newcomer
I posted a question about this in HardOCP forum but i feel i didnt get a real useful response however i thank those who answered, lets see if this forum community can do better.

Im thinking of buying a Athlon 64 gaming system but before i need guess the need of a Athlon 64 PCI Express motherboard.


My conclusion.
The following could be right or wrong so plz correct if im wrong.
My gueess right now is that there is no need for PCI Express instead of AGP 8x in 3d gaming since the performance gain from AGP 4 to 8 is virtually none. Todays PC gaming systems use graphic cards RAM instead of the of AGP bus to process data and since the RAM on the current high end graphic cards are 256 MB i see no need for such a fast bus especially since games like doom 3 wont use more than 128 MB RAM on the graphic card without AA and AF. Even with 4AA and 8AF on it wont probobly need more than 256 MB RAM in resolution 1280*1024 on the graphic cards. In the end if newer games requires more bandwith it will be more RAM on the future graphic cards that will be required instead of a faster graphic bus.

Of course new tech is allways welcome but if PCI Express wont offer any performance gains in 3d gaming i feel that there is no need to wait for a Athlon 64 motherboard with PCI Express. Hopefully PCI Express will be useful for 3d gaming in the long run but if that is from year 2009 then ive already swapped maiboards i couple of times before.


I would like to hear your opinion on how much PCI Express will improve 3d gaming performance since my conclusion above could be TOTALLY wrong.

Pardon my english and thx for any replies.

/Nevyn
 
I believe Ati's upcoming boards, based on the R420 graphics core, will be pci express. Isn't that reason enough to upgrade ;)?
 
AF doesn't affect the amount of memory needed on the graphics card, whereas AA does.

Other than that, my general impression is that the upgrades that have been done to AGP usually only show benefits with GPUs that appear a couple of years down the line - so I would expect the actual performance gap between AGP8x and PCI-Express platforms to be 0-2% for the first couple years after PCI-Express gets common (barring issues with GPUs that ditch AGP support). This will get even smaller as VS/PS3.0, uber-buffers and the rumored PPP(Programmable Primitive Processor) allow geometry data to be generated under user control completely onboard.
 
I think it depends on whether you plan to upgrade incrementally or you buy a whole system after a set period.

If you buy a whole system every 12-24months or longer as most people do then there is no need to worry about PCI-Express.

If you plan to upgrade your graphics card for example in 12 months but want to keep your motherboard, then wait for PCI-Express.

Eventually PCI-Express graphics cards will be cheaper than the AGP versions a bit like what is happening with PCI vs AGP graphics cards. The issue isnt about speed at this time because yes PCI-Express has benefits (otherwise what would be the point of it) but rather your buying/upgrading habits.
 
Tahir wrote :

Eventually PCI-Express graphics cards will be cheaper than the AGP versions a bit like what is happening with PCI vs AGP graphics cards. The issue isnt about speed at this time because yes PCI-Express has benefits (otherwise what would be the point of it) but rather your buying/upgrading habits.


Without making a huge report can someone simply point out where PCI Express will offer big improvements over AGP that will be noticed in reality and not just on a tech-papper.

thx

/Nevyn
 
I think you're asking for the impossible, namely predicting the future.

You can likely come up with as good an answer as anybody else. PCI-Express is faster, better, etc etc.

If that will pay out in reality waits to be seen. Everything else is just speculation, and that pretty much anyone can do.

Tip: when you ask for free advice and don't hear what you'd like to hear, don't complain okay?


*G*
 
Intel always increases the bus speed before it becomes a huge bottleneck so initially I don't expect games to see a benefit. Just look at the various AGP transitions for proof of this.

I agree with Tahir in that waiting depends on how long you want to keep your motherboard for. In my case I bought an AGP4x motherboard when it first came out and it was a good decision because I still use that computer. I'm now planning to upgrade, only this time I'm not waiting for PCI Express because I plan to shorten my upgrade cycle so stay up with mainstream performance.
 
NVIDIA suggests that at least in some cases, up to a 20% graphics performance boost can be achieved by transitioning from AGP to PCI Express. That's according to one of their presentations at Fall IDF, which was helpfully linked (today) by nvnews.net.

The acceleration isn't because PCI Express 16X has double the bandwidth of AGP8x (which it does), but because it uses the bus more efficiently and can avoid conflicts better than AGP/PCI. I'm sure that the extra bandwidth doesn't hurt, but as you pointed out, the transition from AGP 2x->4x->8x was rather uneventful.

The NVIDIA video does a nice job of explaining why AGP was a significant development, though; I found it informative.

As for whether waiting for PCI Express is a good idea... I dunno. The first motherboards are likely to be expensive, and it will be a while before anything other than graphics cards use the bus. IMHO, a year from now would be a good time to join the party.
 
the biggest improvement is, or, will be, the fact that both ways are at the same speed. today, a lot of stuff (that is cheap and often used on consoles, for example), can't be accomplished, because the readback of data from the gpu is essencially pci. very slow.

with pci-express, data can transfer very fast from and to the gpu, from and to the cpu. this means they can share tasks.. it will be possible to do the physics simulation partially on gpu, and possibly some render-process part on the cpu.. how ever. where the cpu lacks the speed of the pipelined hw, we can use the gpu, where the gpu lacks the full programability (even vs/ps3.0 isn't really much compared to a cpu), we can use the cpu.. both can work on the same data..

this, till now, was never possible. it is possible on an xbox, for example (as they share the same memory even:D), as well as other consoles.. gamecube f.e.
and they can accomplish tons of cool stuff with it. espencially post processing of the image gets very extensible..
 
Luminescent said:
I believe Ati's upcoming boards, based on the R420 graphics core, will be pci express. Isn't that reason enough to upgrade ;)?
IIRC R423 is PCIE, R420 is AGP ;)
 
Grall wrote :
Tip: when you ask for free advice and don't hear what you'd like to hear, don't complain okay?


hmm im not complaining but if it seems so im sorry i must have expressed myself badly then.

I want thank everyone who have posted replies.



/Nevyn
 
davepermen said:
the biggest improvement is, or, will be, the fact that both ways are at the same speed. today, a lot of stuff (that is cheap and often used on consoles, for example), can't be accomplished, because the readback of data from the gpu is essencially pci. very slow.

with pci-express, data can transfer very fast from and to the gpu, from and to the cpu. this means they can share tasks.. it will be possible to do the physics simulation partially on gpu, and possibly some render-process part on the cpu.. how ever. where the cpu lacks the speed of the pipelined hw, we can use the gpu, where the gpu lacks the full programability (even vs/ps3.0 isn't really much compared to a cpu), we can use the cpu.. both can work on the same data..

this, till now, was never possible. it is possible on an xbox, for example (as they share the same memory even:D), as well as other consoles.. gamecube f.e.
and they can accomplish tons of cool stuff with it. espencially post processing of the image gets very extensible..
AGP does have a mechanism that allows the GPU to write to system memory at full AGP speed (2 GB/s at 8x), just as fast as reads and fast-writes. However, Nvidia and ATI have so far not bothered to implement this mechanism, forcing the user to rely on slow PCI reads to retrieve data from the graphics board. I suspect that PCI-Express graphics cards will suffer from a similar lack of optimization.
 
davepermen said:
nope, as they can't go back trough some pci-portale.. there is "just" the fast way:D
Unless there is something I have missed with PCI-Express, you might very well still need to initiate every transfer from the CPU side, which is slow - the result might very well be that you pay 1 full roundtrip latency for every word of data you transfer, just like with old PCI. If that happens, you will get a surprisingly small speedup over today's way of doing PCI transfers over AGP.
 
Arjan is right that PCI-Express is no guarantee that feedback will be fast. There's still good opportunities left to clog up the data. :( PCI-Express can't transfer the data faster than the surrounding hardware feeds it.

But at least there's a better chance that it will be well implemented this time than before, since the benefit of feedback has become more and more obvious lately.


I personally would like that the new interfaces will be accompanied with a standardized way to DMA data from a PCIE-1x card straight into a texture on a PCIE-16x card. The obvious use would be to make the gfx+tv cards redundant. No more need to update both in pairs. Even PCIE-1x is fast enough by far to feed a TV stream. (Even though I'd love the functionality of a AIW, I would never by an AIW card.)
 
well, but the fast way is rather enforced.. and even nvidia and those "see" that now.. (watching some pci-express videos..)
 
Basic said:
I personally would like that the new interfaces will be accompanied with a standardized way to DMA data from a PCIE-1x card straight into a texture on a PCIE-16x card. The obvious use would be to make the gfx+tv cards redundant. No more need to update both in pairs. Even PCIE-1x is fast enough by far to feed a TV stream. (Even though I'd love the functionality of a AIW, I would never by an AIW card.)
The DirectDraw API has a standard way of reserving a surface for incoming DMA data from a TV card, and PCI/AGP buses have been allowing such card-to-card DMA transfers for at least a decade. I have a PCI TV card myself (an OLD Hauppauge WinTV/pci) that uses DMA to transfer TV pictures from the TV card to the graphics card; it has survived at least 6 years and 4 graphics card upgrades.
 
personally I'm glad AGP is being upgraded "before it's needed." could you imagine if graphics cards still used the PCI bus?
 
nobie said:
personally I'm glad AGP is being upgraded "before it's needed." could you imagine if graphics cards still used the PCI bus?

Yea, Everyone would be using PowerVR Chips....

*grin*
 
arjan de lumens said:
The DirectDraw API has a standard way of reserving a surface for incoming DMA data from a TV card, and PCI/AGP buses have been allowing such card-to-card DMA transfers for at least a decade. I have a PCI TV card myself (an OLD Hauppauge WinTV/pci) that uses DMA to transfer TV pictures from the TV card to the graphics card; it has survived at least 6 years and 4 graphics card upgrades.

Heh, I must have missed something then. Can it load into a texture? And it should be possible to make the texture triple buffered In case the rendering can't be synchronized to the TV.

I also have an old Hauppauge WinTV/pci that has lasted a few graphics cards, but all attempts to do anything with the video stream myself has turned out slow as molasses. I guess I've been using wrong interface.
 
Back
Top