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

[Mechromancer]

This is a PCI Express speed test that measure CPU to GPU bandwidth, and GPU to CPU bandwidth. It is very stressful on your system so do not be alarmed if your ATI Catalyst driver crashes and recovers to desktop. According to the AMD forums that is to be expected in most cases. Take your screen shot mid-run to get your results.

I started a discussion about this benchmark and the results in Tony's post about HT Link speeds. I'm hoping this program will help us diagnose some of our system problems and performance anomalies. Here is a link to that discussion: http://www.xtremesystems.org/forums/...=224502&page=5

This is my output:


As you can see, I have a bandwidth problem in my GPU to CPU speed. It is far lower than Wex's by about FIVE times! I ran this on the system in my sig (M3A79T- Deluxe based), while he did his on a M4A79 and 4870x2.

Show us your results to compare!

EDIT: This test works far better from the CMD command line, as the window doesn't immediately close out if your system is incredible enough not to go into VPU Recovery due to the stress.

EDIT #2: Along with your results, please post the OS you are running as well. Over at the AMD Developer Forums where they are having the same discussion, they say this program will likely crash on x64 and Vista, but should be stable on XP x32. There is also other interesting info and test results to look at there as well. The program runs in linux too, so I will boot up my OpenSUSE 11.1 if I can get the drivers to work right and post some linux results.

[i found nemo]

i know i have agp, but here u go



how far we've come lol

[Mechromancer]

i know i have agp, but here u go



how far we've come lol

How far indeed!

Wex's GPU to CPU bandwidth was over 4GB/s@~1800mhz HT Link speed which worries the heck out of me.

These were my fastest scores:
Max speeds at CPU Core@3.2 1.35v, NB@2.4 1.25v, HT Link@1800mhz 1.2v, PCIe@124mhz, HT, NB Chipset@1.1v:
Max CPU -> GPU: 5.484 GB/sec [8388608 bytes]
Max GPU -> CPU: 943.053 MB/sec [1048576 bytes]

HT Link@2200mhz 1.2v:
Max CPU -> GPU: 6.417 GB/sec [8388608 bytes]
Max GPU -> CPU: 952.805 MB/sec [33554432 bytes]

If other M3A79T-Dlx users, or other people have scores like mine, while others have much higher, we potentially have found an issue. I can't imagine it could be board design though.

[Pyr0]

interesting, here's a quick run on my sys at stock:

[Smartidiot89]

Indeed interesting Mechromancer!

I've seen your research and I am starting to believe - hey HTLink isn't useless

The real question is, how will the PCI-E speeds affect real-world apps and gameplay? Or is this just another type of benchmark?

[Mechromancer]

Indeed interesting Mechromancer!

I've seen your research and I am starting to believe - hey HTLink isn't useless

The real question is, how will the PCI-E speeds affect real-world apps and gameplay? Or is this just another type of benchmark?

Yes, that is the question! That is a question Tony and others set out to answer in his post that I reference above. This PCIe Speed Test is a tool from the AMD Stream Power Toys site to let us actually see REAL bandwidth between the CPU and GPU. The two factors that seem to affect this bandwidth are the actual PCIe frequency and the HT Link frequency. The affect on bandwidth is very large.

I'm also trying to diagnose a bandwidth issue on my system with this tool. It appears something is configured wrong. I first thought my system wasn't giving all 16 lanes to my GPU, but then I see the CPU-to-GPU speed is fine. The GPU isn't communicating to the system properly however .

Let us know how well your system does with this test.

[i found nemo]

agp @ 76 mhz

[BeepBeep2]

My PC locked up when it got to 268435456 bytes.

PC in my sig gave 2.8GB CPU to GPU up until it froze.
It's Pcie 1 so it should be getting about half of what you guys are getting, but way more than AGP.
EDIT: I ran it again...
http://img32.imageshack.us/img32/5300/pciebench.jpg

Notice the 1.8GB GPU to CPU...

[Daveburt714]

I noticed you have the BIOS pci-e tweak turned on Mech. I had mine off with the clock set to 115 manually and had trouble getting past the 67meg test, turning it down to 105 gets me much further along but still crashes the drivers @ the 512meg test (which I assume is normal).

Considering the difference in clock speeds it seems like our scores are close. I'm also running an M3A79-T (see sig).

I get that wierd "glich" @ the 134meg mark on alot of my runs.

Wez's scores blew me away as well, at first I thought he may be running DDR3 but his sig says 940 so that can't be the case....

If other M3A79T-Dlx users, or other people have scores like mine, while others have much higher, we potentially have found an issue. I can't imagine it could be board design though.

I couldn't agree more. Hopefully we'll see some more results from folks.


PCI-e Clock @ 105

[evilSuRGe]

2 x 3870x2 in crossfire

[Mechromancer]

I've noticed that Wez and evilSuRGe both have X2 GPUs. I'm wondering if the X2s magically put out 3000+MB/s of bandwidth more than regular GPUs. Also all the different scores on these different platforms is making me have a WTF moment right about now....

Something serious is going on here. We need more tests from different users to compare.

@Daveburt714: Oops, his name is Wez not Wex lol. Also, I'm glad to see our scores are pretty closer to each other. It doesn't quite look like an M3A specific issue, but I wonder how the board is configuring the PCIe lanes when multiple PCIe slots are populated.

[Slagathor]

Rig in sig, not overclocked, NB and core at stock, ram at 1066

[Daveburt714]

Rig in sig, not overclocked, NB and core at stock, ran at 1066

Is that 2x 4870's in crossfire?

[Mechromancer]

2 x 3870x2 in crossfire

Can you lower you HT Link Speed to 2Ghz or 1.8Ghz and do a run? I'd like to see if that increases or decreases performance. According to Tony's post, HT link speeds higher than 2Ghz actually hurt graphics performance. Since PCIe bandwidth is affect so much when we change our link speeds, maybe changes in bandwidth are the cause.

[Mechromancer]

Rig in sig, not overclocked, NB and core at stock, ran at 1066

Wow, your system looks like it's running fine, but at PCI Express v1.0 speeds and not 2.0 speeds. Run a GPU-Z and it should tell you if you GPU is running at version 1.0 or 2.0 PCI Express. With DFI, you probably have to manually set it in the BIOS. Let us know what you find.

[Slagathor]

Is that 2x 4870's in crossfire?

Yes

[Slagathor]

Wow, your system looks like it's running fine, but at PCI Express v1.0 speeds and not 2.0 speeds. Run a GPU-Z and it should tell you if you GPU is running at version 1.0 or 2.0 PCI Express. With DFI, you probably have to manually set it in the BIOS. Let us know what you find.

Not sure what I'm looking for.....heh





EDIT:
This is what Everest tell me....

PCI Express Controller
PCI-E 2.0 x16 port #0 In Use @ x16 (ATI RV770 - High Definition Audio Controller, Sapphire Radeon HD 4870 512MB GDDR5 Video Adapter)
PCI-E 2.0 x2 port #1 In Use @ x1 (Marvell Yukon 88E8052 PCI-E ASF Gigabit Ethernet Controller)
PCI-E 2.0 x2 port #2 In Use @ x1 (Silicon Image SiI 3132 SATARaid Controller)
PCI-E 2.0 x16 port #0 In Use @ x16 (ATI Radeon HD 4870 (RV770) Video Adapter, ATI RV770 - High Definition Audio Controller)

[Mechromancer]

Not sure what I'm looking for.....heh





EDIT:
This is what Everest tell me....

PCI Express Controller
PCI-E 2.0 x16 port #0 In Use @ x16 (ATI RV770 - High Definition Audio Controller, Sapphire Radeon HD 4870 512MB GDDR5 Video Adapter)
PCI-E 2.0 x2 port #1 In Use @ x1 (Marvell Yukon 88E8052 PCI-E ASF Gigabit Ethernet Controller)
PCI-E 2.0 x2 port #2 In Use @ x1 (Silicon Image SiI 3132 SATARaid Controller)
PCI-E 2.0 x16 port #0 In Use @ x16 (ATI Radeon HD 4870 (RV770) Video Adapter, ATI RV770 - High Definition Audio Controller)

It looks like you are running at only PCIe 1.0 speeds. Congratulations! You can now unlock a lot of performance. Double check with the latest version of GPU-z though. My output looks like this:



The part highlighted in red is shoing I am running a PCIe v2.0 speeds.

[Slagathor]

I found it, wow......

[Zeus]

Prog ain't working for me, it says can't find aticalrt.dll.

[Mechromancer]

Prog ain't working for me, it says can't find aticalrt.dll.

What version drivers are you using? I think the site says you ahve to have 9.2 or newer drivers. I think most of us are using the latest Catalyst 9.5s.

[Mechromancer]

I found it, wow......

WOW INDEED! Run some 3D benchies and let us know if your scores are a lot higher. It looks like your gaming rig is a lot healthier now.

[Zeus]

ermmm... 9.1...*hides in shame*

OK here goes:



Htt link 2000MHz, CPU 3.6GHz, CPU-NB 2.6GHz, 1x HD4870

Good or bad?

[Mechromancer]

That is effectively the same output as Daveburt's and mine.



For some reason the single chipset graphics cards have a lower GPU to CPU bandwidth than dual chipset cards. We need an ATI engineer to explain this to us. I'm sure we have one or two skulking around the forums...

[Slagathor]

After faster PCIe



Before:



Also I notice frames are far more stable in Fallout 3, I run it maxed at 1600*1200 8xAA 8xAF in ccc. Hardly drop below 60 any more, although max is 75, there must be a cap, but I never looked into it....