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

[MadDias]

phemonII 955 @3.84GHz/2.64GHzNB
2.1GHzHT
2x2GB DDR3 1600
HD4870X2 800/950 oc in ccc, but i don´t know if 3d clocks are used for this test...

[evilSuRGe]

Here are a few runs with HT @ 1830, 2080, 2340
I dont see a big difference on when i change PCIe bus from 124Mhz to 100Mhz as you can see below

[evilSuRGe]

Here's one at PCIe @ 134Mhz
I tried 200Mhz but she crashed
Gonna see how far i can push it.

[evilSuRGe]

heres 2600. Its slow at that speed for me.

[aGeoM]

Hi guys

Well... I did run the PCIeSpeedTest (thanks Mechromancer), can't remember BIOS settings, here are the results:


By ageom at 2009-05-30

[Baam]

Anyone using this with Windows 7? I tried setting it to Vista compatibility mode but still crashes.

[i found nemo]

lol ageom is running win 7

and so am i

runs fine

[Stoner133]

I ran this on my 940 and here are the results, don't know if this is good or bad.


By knuckles6 at 2009-05-30

[pawl]

ive tried to run this under XP64 with catalyst 9.5 and it totally bucktooths my system. i get vpu recover and then i gotta reboot cuase the video is totally phreaked out.

[Jaivan]

Rig in sig.
Running at x8 because of raid card.

[Mechromancer]

ive tried to run this under XP64 with catalyst 9.5 and it totally bucktooths my system. i get vpu recover and then i gotta reboot cuase the video is totally phreaked out.

The driver recovery is normal. It is a VERY intensive little program. When you get to the larger transfer sizes it will appear that your computer has locked. It hasn't though. My mouse and keyboard respond very slowly though. I actually finished a couple of runs yesterday. The whole thing took like 10 or 15 minutes to run.

[XRogerX]

ok i dont know if this is good or now so would someone be so kind to let me know

Thanks Roger

[Lightman]

Interesting thread!

Here is my input (CF):







Rig in the sig!

[Oldguy932]

I ran it on my system and the only anomaly compared to everything else was that with the ht link at 1:1 with the nb at 3072 it got 20 gb/s cpu to gpu for the 128mb test and crashed at the 256mb test. Dropping the ht link down 2 notches on the multi, it passed the 256mb test but no more anomaly. Almost like that one time I got absolutely ridiculous scores in sciencemark.

[i found nemo]

some of you have bad results, makes my agp system look decent :P

[Mechromancer]

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!

[Lightman]

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

[Oldguy932]

So if someone could fool the driver into thinking that a single card was actually dual we could basically get better performance?

And has anyone done tri-xfire or quad xfire and not just two X2 cards?

[Mechromancer]

So many questions here...

I wonder if ATI chipset engineers understand that a single GPU needs this amount of bandwidth, with the assumption being the other PCIe slots may be used by other devices. I personally use a x8 RAID controller and an x1 TV Tuner. A person with a multi-GPU system is the type of person that wants full bandwidth for their GPUs and probably doesn't have the other PCIe slots populated with any other type of device. The system allows the GPUs to have as much bandwidth as they can get in this case.

That explanation is ok, but it still begs the question, why not give each slot what it asks for regardless of the device type? If you have one GPU with no other devices on the PCIe bus, it should get all the bandwidth you can throw at it. If you have a setup like mine, then my 790FX should give my x16, x8, and x1 cards what they're asking for. I will take a look at my BIOS to make sure each individual slot is given what I want it to have (gotta love the M3A79-T Deluxe bios). Modern Intel and AMD chipsets have a set number of PCIe lanes to work with. The slot and peripheral PCIe config on the boards is set up such that you don't run out of lanes. I don't see why my x16 slot can't give full x16 bandwidth bi-directionally as is the PCIe spec. If this is an intentional driver bandwidth limit, I want an explanation as why it HURTS performance. The only reason to limit anything in my XTREME point of view is because it yields diminishing returns, which hopefully is the case.

Still, lets figure out if it's the driver or motherboard that is doing this. Can we get any Intel Core i7 X58 and Core 2 guys with ATI GPUs to run this program? This will let us know if it is a chipset or display driver issue, if those at all.

[Final8ty]

So if someone could fool the driver into thinking that a single card was actually dual we could basically get better performance?

And has anyone done tri-xfire or quad xfire and not just two X2 cards?

Tri Fire. 3x3870 2x3870 gave the same


Single 3870


Tri Fire. 3x3870 bumped up PCIE from 100Mhz to 175Mhz. 1.250v on the PCIE. Ran PCIE at 200Mhz did not improve anything from 175Mhz & i don't feel like pushing to 250Mhz.


VC1 Traffic, Isochronous Flow Control Mode, UnitID Clumping 2/3&B/C, 2x LCLK Mode all on. On all tests.

[pawl]

hmm. tried leaving it to run a while. definate freezing at random intervals. i tried leaving it a few minutes - no luck. i may leave it to churn while i go to sleep. but it stops after a few seconds so i dont hold out much hope for a full run.

it seemed to run better when invoked from cmd directly and not from explorer. also seems to get further with ccc closed.

is anyone running this tool successfully under an x64 system?

my cpu->gpu is about 5gb/s ----- but my gpu->cpu is only geting to about 1.0gb/s

[Lastviking]

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....

Same HT Link?, only change is higher PCI-e?

[Oldguy932]

I noticed that stability of the pcie can actually be determined with everest memory speeds. My system hits hits around 9975-10010 in the memory read test when everything is stable at the speeds. It turns out that with pcie at 140mhz at the ht link up at 3072 its not stable with 1.2v on the ht. I knew that because the memory reads were in the 9600s and 9700s meaning something wasn't right. With the voltage bumped up to 1.28v it works fine now and memory reads are back where they should be.

The ht link needing more voltage goes back to the other thread about when I asked why not have the ht link higher since it should bring higher performance. I'm guessing mine didn't need voltage till a while later because the cpu is on phase, but air systems need voltage sooner.

[Mechromancer]

Tri Fire. 3x3870 2x3870 gave the same


Single 3870


Tri Fire. 3x3870 bumped up PCIE from 100Mhz to 175Mhz. 1.250v on the PCIE. Ran PCIE at 200Mhz did not improve anything from 175Mhz & i don't feel like pushing to 250Mhz.


VC1 Traffic, Isochronous Flow Control Mode, UnitID Clumping 2/3&B/C, 2x LCLK Mode all on.

Are all of those test with the Isochronous Flow Control options enabled?

hmm. tried leaving it to run a while. definate freezing at random intervals. i tried leaving it a few minutes - no luck. i may leave it to churn while i go to sleep. but it stops after a few seconds so i dont hold out much hope for a full run.

it seemed to run better when invoked from cmd directly and not from explorer. also seems to get further with ccc closed.

is anyone running this tool successfully under an x64 system?

my cpu->gpu is about 5gb/s ----- but my gpu->cpu is only geting to about 1.0gb/s

I only run x64 baby! x64 performance is the only performance that matters LOL.

I noticed that stability of the pcie can actually be determined with everest memory speeds. My system hits hits around 9975-10010 in the memory read test when everything is stable at the speeds. It turns out that with pcie at 140mhz at the ht link up at 3072 its not stable with 1.2v on the ht. I knew that because the memory reads were in the 9600s and 9700s meaning something wasn't right. With the voltage bumped up to 1.28v it works fine now and memory reads are back where they should be.

The ht link needing more voltage goes back to the other thread about when I asked why not have the ht link higher since it should bring higher performance. I'm guessing mine didn't need voltage till a while later because the cpu is on phase, but air systems need voltage sooner.

This reminds me of an interesting point I learned about HyperTransport. HT uses LVDS, Low Voltage Differential Signaling (READ HERE) to signal at high speeds, while minimizing signal noise. Because of how HT works electronically, I think there may be a bigger relationship between voltage and frequency than just what voltage you need to make it work. The wrong voltage may impair the actual communication and degrade performance at certain frequencies.

With that being said, the spec says HT is supposed to work all the way up to 2600Mhz at 1.2v so . Mine cannot!

[Final8ty]

Are all of those test with the Isochronous Flow Control options enabled?

On all tests yes.