So why exactly did ATI move to PCIe again?

R

Rolf N

Recurring Membmare
Veteran
That's what they claim.
"ATI's true PCI Express graphics design provide [sic] up to TWICE the bandwidth of AGP8x. ATI's true PCI Express graphics solution ensures full bandwidth is available in both upstream and downstream directions."

But don't believe a word. I just got the numbers in. The guy obviously got ripped off by ASUS (they "specify" a 128 bit memory interface but deliver only a 64 bit one), but that's beside the point.

My point here and now is that readback speed of ATI's x1300 PCI Express stuff is 200MB/s. That's 10% more than any Geforce <=FX on AGP. It's one third of any Geforce 6 or any Wildcat VP series on AGP. So they forfeit the single marketable-to-users benefit of PCI Express over AGP.

My initial thought was just "Fuck you ATI". That move must have really been worth it. Currently it just boggles my mind.
 
ATI's readback sucks under current OpenGL drivers, its more optimised for DX. Serious Magic readback test approached double their AGP performance with X800 XT PCIe. The Serious Magic performance has also increased since then and is a little improved with X1800.
 
zeckensack said:
But don't believe a word. I just got the numbers in. The guy obviously got ripped off by ASUS (they "specify" a 128 bit memory interface but deliver only a 64 bit one), but that's beside the point.
Click to expand...
Not really, I had feeling that that that card doesn't have an 128bit Memory Interface cause no low profile card usually has...

Even worse is that this ASUS X1300 can't be used as a low profile and that really sucks..

@Dave
I've test that programme later, in about 10 hours or so.

So far about 200MB/sec on my 'office computer' with an Radeon Xpress 200, Intel edition.
 
Last edited by a moderator:
With Vsync disabled reference X1300 gives about 657MB/s with the Serous Magic test (DX) at 1024x768. Stanfords GPU Bench readback tests (OGL) give between 260-292MB/s.
 
Stefan Payne said:
Not really, I had feeling that that that card doesn't have an 128bit Memory Interface cause no low profile card usually has...
Click to expand...

I got confirmation that this Asus graphic board has a 64 bit memory bus. Asus will correct the spec on their webpage.
 
Dave Baumann said:
With Vsync disabled reference X1300 gives about 657MB/s with the Serous Magic test (DX) at 1024x768. Stanfords GPU Bench readback tests (OGL) give between 260-292MB/s.
Click to expand...
I've got about 615MB/sec, on my 'good' nF4 board.
But why is the Xpress 200 thing so slow? :|

@Tridam
That sounds not soo bad.
But why built a low profile card that can't be used as a low profile card?
 
Stefan Payne said:
I've got about 615MB/sec, on my 'good' nF4 board.
But why is the Xpress 200 thing so slow? :|
Click to expand...
Sorry? That 200MB/s figure was with the Serious Magic test? If so it might be something to do with the fact that its an Intel platform and the processing thats going on - I don't exactly know what the tests are doing but I would guess that the data is being rendered to the frambuffer, that is being read back through the GPU, over the PCI Express interface and into system memory, so its probably already very dependant on other elements, and if it has to do a roundtrip via the CPU before ending up in system RAM then that could be why its slower. (I'm using the AMD XPRESS 200, BTW)
 
Also ensure that Z sync is disabled with the Serious Magic test - it doesn't override VSync automatically and locks to 25FPS.
 
Dave Baumann said:
ATI's readback sucks under current OpenGL drivers, its more optimised for DX. Serious Magic readback test approached double their AGP performance with X800 XT PCIe. The Serious Magic performance has also increased since then and is a little improved with X1800.
Click to expand...
Thanks, that's mildly comforting. There seems to be a gap between DirectX Graphics and OpenGL in terms of readback speed, and while that gap was small in previous hardware generations, it appears that it has widened significantly with the latest chips.

I.e. I get 100MB/s back from my Radeon 9600SE/AGP* in OpenGL, and 125MB/s with the Serious Magic test. That's at least the same ballpark. I couldn't make out from the description whether the test reads the actual backbuffer or a more optimally formatted render target (parallel to OpenGL's "rectangle" textures, which are flat in contrast to standard 2D textures and frame buffers which are usually swizzled in some way). If it's the latter, it might explain the small difference in this case.

*don't ask; I had bruntier AGP cards but sold them
 
Stefan Payne said:
I've got about 615MB/sec, on my 'good' nF4 board.
But why is the Xpress 200 thing so slow? :|
Click to expand...
You mean it's actually not the card but the PCIe implementation on the chipset/board?
Or did you use Serious Magic for these numbers and OpenGL still stinks?

Please do tell. I'd actually love to be wrong in this case.
 
Does annyone have an Xpress 200 Intel Edition (exept me)??

Somehow that thing sux, with Serious Magic I got about 150MB/Sec :O :O

I think I've should test the P4C instead of the P4B and PC3200 RAM...
 
Stefan Payne said:
Maybe its the PC2100 RAM in that thing...

If I think about it, I test it with PC3200 RAM.
Click to expand...
Stefan Payne said:
Does annyone have an Xpress 200 Intel Edition (exept me)??

Somehow that thing sux, with Serious Magic I got about 150MB/Sec :O :O

I think I've should test the P4C instead of the P4B and PC3200 RAM...
Click to expand...
You're more than an order of magnitude away from any possible system memory bandwidth limitations.
I very much doubt that readback performance will increase at all with faster memory.
(general performance will improve, of course)
 
zeckensack said:
You're more than an order of magnitude away from any possible system memory bandwidth limitations.
I very much doubt that readback performance will increase at all with faster memory.
(general performance will improve, of course)
Click to expand...
You'll never know 4 sure unless you've tested it...

In this case your right but with my PC3200 RAM and the P4C (both 2.4GHz) the Performance with this terrebly loud x1300 increases by about 50MB/sec with the Serious Magic thing.
 
What FPS are you getting with the Serious Magic test (I get the impression that you might have to force VSync off).
 
Dave Baumann said:
What FPS are you getting with the Serious Magic test (I get the impression that you might have to force VSync off).
Click to expand...
about 62fps by ~215MB/sec.

VSync is off, I've tested with 75Hz...
Still the x1300 on my ASUS P4RD1-MX, this time my 2.4GHz P4C (200MHz FSB, 200MHz RAM).
 
GPUBench has no direct connection to the folding client. However, good overall GPUBench scores for instruction issue, input/output bandwidth, and latency hiding should lead to good performance in most GPGPU applications. Few GPGPU apps are readback limited.
 
You must log in or register to reply here.
Back
Top