ATI PCIe Speed Test v0.1 CPU-GPU/GPU-CPU bandwidth test

[Mechromancer]

As a response to us over in the AMD forum, I would like to see how ATI GPU PCIe performance on Intel Core 2 or Core i7 chipsets perform. We've found some interesting results over in our thread. Refer to it for more information: http://www.xtremesystems.org/forums/...d.php?t=225823

Give it a try on your single and multi-GPU systems. On the AMD side, we see there is a difference in actual GPU to CPU bandwidth speed between single and multi-GPU on even the same systems as if the PCIe link speed is increased for multi-GPUs.

This test is also useful to visually see the effect overclocking has on bandwidth. Fyi, it is very stressful and your system may appear to hang due to the stress. Don't worry, it hasn't crashed your system. Also be aware that your driver may initiate a VPU Recovery due to the stress. This is stated to be normal by the developer. This program is from the ATI Stream Power Toys site: http://developer.amd.com/gpu/ATISTRE...s/default.aspx

ENJOY!

[Mechromancer]

Bump. Show your PCIe bus some love!

[Vienna]

Intel System, ATi card, Windows 7 RC1 Build 7100, x64, PCIe 2.0

[Shalamay]

e8600 @ 4.0ghz and 4870x2 vista 64

[OC Nub]

e8400 4.4ghz HD4890

[Mechromancer]

e8400 4.4ghz HD4890

Is that a SINGLE 4890?! If so, you aren't having the issues that single 4870 GPU users are having. Your GPU to CPU speeds match your CPU to GPU for the most part (which is the way an x16 link should work). Interesting indeed!

[OC Nub]

Yes that's a single, put the second card in just now, will test it. I will be running at 8x with cf, p45 not an x38/48.

[OC Nub]

This is with cf 4890's e8400 4.4ghz. Had to take the ss before the test ended otherwise the program would stop responding.

[SoulsCollective]

I'm trying to run it on my 4870X2/i7 920, but the window closes as soon as it finishes the 1GB test on the second core/mem pair - are results saved to anywhere?

[OC Nub]

I'm trying to run it on my 4870X2/i7 920, but the window closes as soon as it finishes the 1GB test on the second core/mem pair - are results saved to anywhere?

My screen flickered with a single card and the program stopped but was able to take a ss. With cf I ran it a couple times to see where it crapped out to know when to hit print screen, otherwise it would stop responding and close. I have no idea if it keeps a log file.

[SoulsCollective]

My screen flickered with a single card and the program stopped but was able to take a ss. With cf I ran it a couple times to see where it crapped out to know when to hit print screen, otherwise it would stop responding and close. I have no idea if it keeps a log file.

See, it runs perfectly for me, no crashes, flickering, anything, and can test up to the full 1GB VRAM. It slows down a lot near the end, so I could take a screenie before it finishes, but that strikes me as a bit pointless.

[Mechromancer]

See, it runs perfectly for me, no crashes, flickering, anything, and can test up to the full 1GB VRAM. It slows down a lot near the end, so I could take a screenie before it finishes, but that strikes me as a bit pointless.

Unfortunately it doesn't appear to save a log file anywhere; however, if you take a look at the readme, I think there is a way. The best way to save your results is just to take a quick screen shot. Usually you'll see your peak CPU to GPU and GPU to CPU speeds well before the end of the test.

I'm also interested in the fact that your system is completely stable with this test. Is your GPU a non-reference design by chance? What is it's make and model? I only run this test completely through maybe 3 times out of 10 without a VPU Recovery.

[Vienna]

Unfortunately it doesn't appear to save a log file anywhere; however, if you take a look at the readme, I think there is a way. The best way to save your results is just to take a quick screen shot. Usually you'll see your peak CPU to GPU and GPU to CPU speeds well before the end of the test.

I'm also interested in the fact that your system is completely stable with this test. Is your GPU a non-reference design by chance? What is it's make and model? I only run this test completely through maybe 3 times out of 10 without a VPU Recovery.

My system was completely stable, and I was/and always am crunching away on BONIC all the time. IO ran the test a second time with BONIC stopped/paused, and the results were the same. Stock, reference ATi card.

[003]

Doesn't work on nvidia cards.

[SoulsCollective]

Unfortunately it doesn't appear to save a log file anywhere; however, if you take a look at the readme, I think there is a way. The best way to save your results is just to take a quick screen shot. Usually you'll see your peak CPU to GPU and GPU to CPU speeds well before the end of the test.

I'm also interested in the fact that your system is completely stable with this test. Is your GPU a non-reference design by chance? What is it's make and model? I only run this test completely through maybe 3 times out of 10 without a VPU Recovery.

I'll run it again, then, and stop at the 1GB test on the second GPU.

The card is a reference design HIS 4870X2, currently at stock 3D speeds, but with 2D clocks dropped to 300MHz for both core and mem for both cores, using Win7 RC and Cat 9.5.

Edit: Here we go:

[gurusan]

E3120 @ 4.05GHz, 4890 @ 940/1000

[Mechromancer]

E3120 @ 4.05GHz, 4890 @ 940/1000

Notice how single 4890 cards have much much higher GPU to CPU communication speeds than single 4870s. The "problem" with the single 4870 that both AMD and Intel users are having must be a driver issue. In CF configurations, the 4870 communicates at full x16 speed back to the CPU.

Damn, I want my bandwidth!

[Mechromancer]

FROM THE AMD FORUM:

Wow...

Crossfire and X2 systems definitely manage system bandwidth much better than single chipset GPU systems. I still fail to understand how the CPU to GPU and GPU to CPU bandwidths aren't close to exactly the same on single GPU systems. My only guess is the GPU doesn't need to send as much information back to the CPU to function, so this speed is artificially limited. CF and X2 systems must need a lot of GPU to CPU bandwidth to coordinate their efforts. This is just a hypothesis/guess after seeing all these scores. Lets get an ATI chipset engineer to explain what we're seeing. Somebody on XS has to know one!

I've came to the same conclusion while thinking about it!
My guess is that as soon as multi-GPU is detected some routines in driver are enabled to handle increased bandwidth between CPU and GPU.

EDIT:
Some proof - this is what I get after simply disabling CF in CCC

Lightman here simply disabled his CF in CCC and he got these sub 1GB/s GPU to CPU numbers. With CPU enabled that number is 5GB/s. Can anybody think of any other explanation other than a driver issue?

[Pyr0]

i7 920 CPU running at 3.6 with a 4870X2, pcie 102

[ahmad]

What the hell..

I thought my oc was stable, reduced it, and this thing still wont go through all the way to 1073741824 bytes test. It does the 536870812 one, and I get the "display driver recovered" message.

How many people have ran this test on C2D and had it go all the way through (where the window closes)??

Vista x64 4870, 9.5 cats.

[Mechromancer]

What the hell..

I thought my oc was stable, reduced it, and this thing still wont go through all the way to 1073741824 bytes test. It does the 536870812 one, and I get the "display driver recovered" message.

How many people have ran this test on C2D and had it go all the way through (where the window closes)??

Vista x64 4870, 9.5 cats.

Since the test stresses the system so much, it is actually normal for the driver to go into recovery mode. This is also per the AMD Forum thread about this test.

[nsegative]

My result

[nsegative]

[SirDaemon]

4870 1GB @ CF
PCIe - 101

[fellix_bg]

• Radeon HD 4890 (reference board);
• Intel Q9450 @ 3800MHz (8*475MHz);
• DFI LanParty P45-T2RS;
• 4GB DDR2 @ 1141MHz;
• Vista x86, SP2;