ATI Radeon X800 XT (PCIe) Review

[Dave Baumann]

http://www.beyond3d.com/siteimages/b3dsmall.gif (http://www.beyond3d.com/reviews/ati/r423/)Prior to the release of the PCI Express platform ATI staked a lot on the transition. Betting that, once introduced by Intel, the OEM’s would be quick to adopt the new bus they sought to align their entire generation of products around the platform, lining up a full range of chips to natively support the platform. While ATI have only just released their latest, and last, high end AGP chip with Radeon X800’s R420, they simultaneously developed R423, which is ostensibly the same as R420 with the interface replaced by a x16 PCI Express interface. Here we take a look at the performance of one of the incarnations of R423 with ATI’s Radeon X800 XT.<blockquote>"Comparing R423 to R420 we see that all the main characteristics of the two chips - process, complexity, configuration - are identical and this is indeed the case. The only difference between the two new high end ATI chips is of course the interface difference, with R420 featuring an AGP8X host interconnect bus interface, whilst R423 replaces this for a 16 lane PCI Express interface."</blockquote>Read the full review here (http://www.beyond3d.com/reviews/ati/r423/)

[madshi]

*Finally* a review with some good AGP vs PCI Express tests - thank you!

Dave, if you've fun you can give our little in house screenshot benchmark a try. It tests how fast our own remote control software can grab screenshots out of the graphics card. We use it to find out which graphics driver to use for which graphics card and which display mode is the best for performance. Maybe it can also show some differences between AGP and PCI Express?

http://madshi.net/videoBench.zip

The interface is in German, but it's so simple you can't do anything wrong. When starting it a box appears, just close it and wait for a few seconds without doing anything (not moving the mouse etc). Then the results should be shown.

[Anonymous]

well... maybe they will now learn that people will NOT upgrade their motherboards, processors and memory just to be able to get a new card on new interface to get a combination that runs slower than the previous...

[Mendel]

well... maybe they will now learn that people will NOT upgrade their motherboards, processors and memory just to be able to get a new card on new interface to get a combination that runs slower than the previous...

my comment, whoopsie about not logging in...

[Sxotty]

People buying from dell don't care though mendel and that is what the card is aimed at. That is why the specs are lowered, it is like the 9700 that dell offered with lower specs than a 9700 pro.

[Evildeus]

This chart seems mess up. You should redo it (XT/XT PE lines)

X800 XT % Difference 640x480 800x600 1024x768 1280x960 1600x1200
X800 XT PE 9.6% 10.1% 9.8% 9.5% -5.4%
X800 XT 500/500 10.5% 10.0% 10.7% 9.9% 1.5%
X800 PRO 9.3% 10.1% 10.6% 13.2% 17.4%
http://www.beyond3d.com/reviews/ati/r423/index.php?p=7

[Dave Baumann]

I don't see what you are getting at?

[Evildeus]

*edit* I'm wrond, just the difference is not the same than previous ones.

[Dave Baumann]

The % difference numbers are all in relation to the PCIe XT - considering its on the P4EE platform and the other only on a Northwood platform thats the difference we'd expect as this test is really a CPU performance test (unfortunatly).

[Evildeus]

Yeah sorry, i was looking once more at the setting. Bot really helpfull unfortunately.

BTW why is there a drop @ 800*600 on the Pixel Shader - PS1.1 test of the PCIe card? A typo?

[Rodéric]

http://www.beyond3d.com/reviews/ati/r423/index.php?p=4
just after the two first tables there's a typo:
"With only a 20MHz difference to the X80 XT" missing a "0"


You should, IMO, add a note somewhere to explain what the label "X800 XT % Difference" means ^^
=> Relative Difference (in %) between the Reference Card and another Card. (How the reference card compares to it)
One (me) could get it as "Relative Difference (in %) between the other Card and the Reference Card" (How it compares to the reference card)


Sometimes you refer to the card has a "X800 XT" others as a "X800 XT PCIe", it would have been better to stick with one naming convention.

[Evildeus]

Sometimes you refer to the card has a "X800 XT" others as a "X800 XT PCIe", it would have been better to stick with one naming convention.I agree on this part (hence my first erronous post :oops: )

[Geo]

Mem speed reported as "500mhz GDDR-3 (1.1GHhz Effective)" Err. . .Also, that same table says "Radeon X800 XT Platinum Edition" when everything else about the article suggests just XT. Or I'm silly. One of the two.

[Anonymous]

Kind of makes the whole review questionable.

[Dave Baumann]

Mem speed reported as "500mhz GDDR-3 (1.1GHhz Effective)" Err. . .Also, that same table says "Radeon X800 XT Platinum Edition" when everything else about the article suggests just XT. Or I'm silly. One of the two.

Updated. Copy and paste issue.

Kind of makes the whole review questionable.

How so? If you read the review it would be fairly obvious what was being tested.

[Anonymous]

how come you didnt test using the same cpu and mobo. these test are hard to rely on because of different cpu's and mobos. to see the real difference wouldnt you have to test using the same test boards.

[Anonymous]

ah whoops i forgot the motherboards either only have an agp slot or pci slots.. grrr

[Dave Baumann]

how come you didnt test using the same cpu and mobo. these test are hard to rely on because of different cpu's and mobos. to see the real difference wouldnt you have to test using the same test boards.

http://www.beyond3d.com/reviews/ati/r423/index.php?p=3

Unfortunately we are not going to be able to get a like for like comparison between the PCI Express X800 XT and the AGP X800 XT Platinum Edition due to the differences between the two platforms. The highest performance Intel CPU available to us for the AGP platform is a Northwood based Pentium 4 3.4GHz, whilst the PCI Express platform requires LGA775 CPU's and only Pentium 4 Extreme Editions (which are similar to Northwood's, but with a much larger cache) or Prescott's (newer P4 cores based on Intel's 90nm process, with numerous pipeline differences and a larger cache) have been transferred across, hence we are using a 3.4GHz Pentium 4 Extreme Edition.

Not ideal, but we have to do the best with what we have.

[Temporary Name]

Great work!

We get the expected outcome (PCI-E 16x is of minimal importance to most gaming applications [the primary usage scenario of high-end video cards], but offers future-proofing), and the somewhat-unexpected outcome (PCI-E 16x offers immediate real-world benefits in specialized, bus-intensive applications).

Anyone pooh-poohing PCI-E as some sort of Intel conspiracy should remember that VIA, SiS, and NVIDIA will all be offering PCI-E-centric chipsets in 2-3 months.

[Anonymous]

Good review. Nice to see a real comparison of PCIx and AGP where it matters. Pity the CPU skewed the gaming results, although it seemed pretty conclusive that PCI Express changes nothing for games. It was also interesting to note that even old games are CPU bound with the current cards.

Seeing the speed of PCI Express, I wonder how cards with slow memory will fair if all their texturing was done from off-card memory. I remember that the Rage Pro had better performance when texturing from AGP memory than on-card RAM. I wonder if it'd be possible to test this for one of the slow X300 cards. For the 64-bit version at least, it's bound to be faster to texture from main memory, leaving the on-board RAM for handling the frame buffer. One could then also make do then with 32MB cards at the low end.

[Dave Baumann]

Dave, if you've fun you can give our little in house screenshot benchmark a try. It tests how fast our own remote control software can grab screenshots out of the graphics card. We use it to find out which graphics driver to use for which graphics card and which display mode is the best for performance. Maybe it can also show some differences between AGP and PCI Express?

http://madshi.net/videoBench.zip

I don't have an X800 on me at the moment. Hopefully I should be getting one back sooner or later. In the meantime I do have the low end FireGL PCIe board, V3100 (basically an X300 with FireGL drivers) and PCIe 6800 GT to test. It'll be a little while before the PCIe board goes back into the test platform though.

[Dave Baumann]

We get the expected outcome (PCI-E 16x is of minimal importance to most gaming applications [the primary usage scenario of high-end video cards], but offers future-proofing), and the somewhat-unexpected outcome (PCI-E 16x offers immediate real-world benefits in specialized, bus-intensive applications).

I would say that both situations were kinda expected though.

[Dave Baumann]

Good review. Nice to see a real comparison of PCIx and AGP where it matters. Pity the CPU skewed the gaming results, although it seemed pretty conclusive that PCI Express changes nothing for games. It was also interesting to note that even old games are CPU bound with the current cards.

Well, it seems that game that use a lot of render to texture and/or normal maps may already be spilling over 256MB frame buffers so there are could be some cases already. TR:AoD seemed to show this, although I was annoyed that FSAA all of a sudden didn't want to work on the AGP boards as I would have liked to have investigated the performance more with larger frame buffers.

Seeing the speed of PCI Express, I wonder how cards with slow memory will fair if all their texturing was done from off-card memory. I remember that the Rage Pro had better performance when texturing from AGP memory than on-card RAM. I wonder if it'd be possible to test this for one of the slow X300 cards. For the 64-bit version at least, it's bound to be faster to texture from main memory, leaving the on-board RAM for handling the frame buffer. One could then also make do then with 32MB cards at the low end.

Interesting thought. I suspect that latencies (Graphics->Northbridge->system RAM) may be one issue and the general lack of being able to reach the theoreticl bandwidth of PCIe would be an issue. Remember, the system bandwidth was topping out at about 4GB/s, so I suspect that with current RAM speeds the difference may be marginal even on the 64-bit cards.

[madshi]

I don't have an X800 on me at the moment. Hopefully I should be getting one back sooner or later. In the meantime I do have the low end FireGL PCIe board, V3100 (basically an X300 with FireGL drivers) and PCIe 6800 GT to test. It'll be a little while before the PCIe board goes back into the test platform though.
No problem at all. You can do a quick bench when it fits into your plans, otherwise don't worry. I'm not even sure whether the bench will show differences between AGP and PCI Express. I think it should, but who knows...

[Tahir2]

A nice review Dave.. to all those that say PCIe does nothing for games and gamers:
Here we see that while the texture requirements are kept low all the boards are performing to their theoretical fill-rate differences, but at texture sizes of 1024x1024, which will certainly require some texture addressing over the bus, the PCIe X800 XT manages to outperform the AGP boards by a fairly significant margin.

More and more games in the near future will be designed to take advantage of the extra bandwidth supplied by PCIe or another way of putting it is that as we get higher resolution textures and more realistic graphics PCIe will show its strengths in a natural way, whether the developer intends for it or not.

PCIe is definitely a move forward for the gaming community and all the naysayers need to realise that PCIe will become the defacto standard in a short period of time due to the way Intel has been so aggressive with supporting it. LGA775 motherboards with Intel chipsets currently come with PCIe only and I hope Intel do not backtrack on this move.

Nice review Dave.

[ET]

PCIe is definitely a move forward for the gaming community and all the naysayers need to realise that PCIe will become the defacto standard in a short period of time due to the way Intel has been so aggressive with supporting it.

I don't think that the nay sayers argue the inevitability of switching to PCIe. The question is whether it's a move forward for gaming is a good one. I mean, sure, it's a move forward like AGPx8 was over x4, but it doesn't feel like having it will suddenly revolutionise gaming.

That said, going forward it could potentially allow more interaction between CPU and GPU (because reading back from the GPU is much faster with PCIe), which might prove useful for the future of 3D and general processing on the GPU.

[ET]

Interesting thought. I suspect that latencies (Graphics->Northbridge->system RAM) may be one issue and the general lack of being able to reach the theoreticl bandwidth of PCIe would be an issue. Remember, the system bandwidth was topping out at about 4GB/s, so I suspect that with current RAM speeds the difference may be marginal even on the 64-bit cards.
I agree that the performance different will likely not be great (it wasn't for the Rage Pro, either), but if there is one, that'd be interesting, and potentially useful for gaining some extra speed (which is always welcome, even if it's 10% or 5%). Regarding the bandwidth, remember that the back buffer and Z-stencil will still be on the card, so the bandwidth will be distributed between card memory and PCIe, which gives this scheme a little more speed than before. IIRC the weight of AA is currently on the bandwidth side, with the pixel shader running just once for all the samples of a pixel, so giving almost the entire bandwidth to the frame buffer could really help.