How to test for PCI-Express bottleneck?

T

TheForumTroll

Weaksauce
Joined
Oct 11, 2009
Messages
105
Hello experts :)

I'm running a brand spanking new GPU on a very dusty motherboard. Is it possible somehow to see the PCI-Express bus utilization and if it is bottlenecking the GPU? I know you can google some tests, opinions and theoretical numbers but I'd like to be able to see/test it myself if possible.
 
gpu -z will test the PCI Express output the card is getting as in 1.0/2.0/3.0/4.0 as to kick start the 3D engine of the video card .

 
Last edited:
TheForumTroll said:
It only shows which PCI Express the card is using though, right? No way to see how much it is using of the available bandwidth.



Sorry, I'm not buying 3D mark pro just for this.
Click to expand...
No, it shows how many lanes and what protocol are being used when the 3D engine is put under stress.

You don't need the Pro version to get the 3D Mark PCI-E Feature Test, just the Advanced version. I bought it on Steam for $10 during a sale one year.

You still haven't answered the question of what video card and motherboard. We can provide more answers if you provide this information.
 
Oh, sorry I misread the comparison of 3dmark versions then. I'll go have another look.

I haven't provided any hardware info because I wanted to be able to test it myself since it is for more than testing one PC. I was only looking to learn :happy:
 
Fire Strike will tell what the gpu is doing as if it below standard as they have a base line the cards should score as where your at .
 
The only thing you can do in this case is run a free benchmark and compare your GPU score to the results of the same card in another machine at the same GPU / Mem frequencies. Even firestrike will vary until you find a card running at the exact freqs (and only look at graphics score)

Now, an RTX 2080Ti barely saturates PCI-E 3.0 x8, which is theoretically a little bit slower than PCI-E 2.0 x16

I'm assuming you're using the P67 board in your sig, so I'm willing to bet if you have any bottleneck it would be negligible unless you're running a top end GPU OR SLI (SLI would be seriously bottle-necked on a modern high end card). Realistically the CPU and memory is always going to be the bottleneck in this situation long before you max out PCI-E 2.0 x16 IMO.

 
Yeah, unless you have a system you can compare to without a bottleneck (both CPU and connection speed), you're going to have to compare to someone online.

Or you can just check the myriad tests already done on major websites, and compare you system's PCIe connection speed to that.

RX 5700 XT PCIe 4.0 test (the 1080p tests are what matter):

https://www.techpowerup.com/review/pci-express-4-0-performance-scaling-radeon-rx-5700-xt/23.html

I wouldn't worry about bandwidth as long as you are still running PCIe 2.0 at x16. I doubt any system recent enough to power a modern graphics card will have anything less than 2.0. a 2% difference between 2.0 and 4.0 is lost in the noise!

The RTX 2080 Ti is SLIGHTLY limited by PCIe 2.0 x16 (but it's still under 5%). As long as you have x16 PCIe 3.0, you're good for the next five years (and with PCIe 4.0, good for ten)!

The tests require low resolution, because the geometry data from so many frames adds way more bandwidth. You should also make sure your CPU can keep up at those lower resolutions.
 
Last edited:
GPUz shows PCIe bus utilization via sensor tab.

This is a fresh, just now, bench of Division 2. The bus interface high was 3%. I have seen it higher before but never much above ~8%-10% with a game, IIRC. Loads for other tasks such as computation are higher. Anyways, Bus Interface Load, lower portion of the window.

 
Last edited by a moderator:
Thank you everyone (y)

Someone mentioned my Sabertooth P67 so I just made a test with it combined with a ROG 5700 XT. I'm not worried by bottlenecks anymore and can tell there is no reason to test other systems for now. But the most important thing is I got a better understanding f how to test these things which was what I was looking for. Thanks! :notworthy:
 
Shikami said:
GPUz shows PCIe bus utilization via sensor tab.

This is a fresh, just now, bench of Division 2. The bus interface high was 3%. I have seen it higher before but never much above ~8%-10% with a game, IIRC. Loads for other tasks such as computation are higher. Anyways, Bus Interface Load, lower portion of the window.

Click to expand...


What GPU-Z shows likely depends on the video card/motherboard/chipset/etc. It does not show Bus Interface Load for my RX570.

GPU-Z-RX570.jpg


This is with all active sensors showing. Just a heads up in case someone else is wondering why it doesn't show up.
 
What GPU-Z shows likely depends on the video card/motherboard/chipset/etc. It does not show Bus Interface Load for my RX570.
Click to expand...


You might be able to select, or not. Options upper right (3 bars), sensors tab, and see if you can select "Bus Interface Load" to be monitored.
 
Last edited by a moderator:
Shikami said:
You might be able to select, or not. Options upper right (3 bars), sensors tab, and see if you can select "Bus Interface Load" to be monitored.
Click to expand...

It's not an option. The screenshot is with all active sensors it has available. There's no other sensors available in the options.
 
The best results would come from a synthetic benchmark that purposely streamed non cacheable, non compressible data from system memory, I would think, to stress the PCIe bus to the max. Surprisingly I couldn’t find something outside of 3dmarks new test in a few minutes of google searching.
A lot of people have tested gpu scaling using more real world loads like gaming on multiple monitors but the issue is most of the time games are not pulling on the PCIe bus that much and use cached data on the GPU to avoid stuttering. The more RAM a card has the harder it is to test without a synthetic test.
 
sethk said:
The best results would come from a synthetic benchmark that purposely streamed non cacheable, non compressible data from system memory, I would think, to stress the PCIe bus to the max. Surprisingly I couldn’t find something outside of 3dmarks new test in a few minutes of google searching.
A lot of people have tested gpu scaling using more real world loads like gaming on multiple monitors but the issue is most of the time games are not pulling on the PCIe bus that much and use cached data on the GPU to avoid stuttering. The more RAM a card has the harder it is to test without a synthetic test.
Click to expand...


And yet, for most users gaming is the highest-load they can come up with. So unless you're using it for compute/game development, it feels like a pointless test set.
 
You must log in or register to reply here.
Back
Top