PCIe Frequency - helping your GPU OC

[wolfram23]

Hi everyone,

I just spent some time fooling around with my PCIe frequency, I wanted to see how it might affect performance. Default PCIe frequency is 100, so I put it up by 2mhz increments. At 110mhz, it still seemed totally stable so I made 5mhz jumps. At 130, I decided to make a big jump to 140mhz where I ran into problems. My audio card stopped working as did my network connection (not sure how that relates to PCIe, but there you have it). I rebooted, dropped it down. Got it working ok at 135mhz, but youtube videos froze without loading so I dropped it once more to 130mhz, and it worked fine.

First of all, here's my BIOS settings - 200x20 4ghz i5 750, 1600mhz CL8 RAM, etc.

In windows, you can see CPUZ, GPUZ, and what the Afterburner graph looked like. You'll see my 5850s are at 960mhz core, 1200 memory.

Finally, the imporant part - how the PCIe frequency helped.
First of all, we'll compare at 0xAA 1920x1080 full screen FurMark test.

100mhz

130mhz

Scores go from 20682 to 21054. Only a 1.8% increase.

Now with 4xAA
100mhz

130mhz

Scores go from 9658 to 9758. This gives a 1% increase again.

This, to me isn't quite satisfying enough but it might be a limitation of FurMark, so I'll try a couple more tests - AvP Benchmark and Dirt 2 Benchmark. Both are DX11, so CPU bottlenecking shouldn't be an issue (DX11 supports multithreading)

First up, Dirt 2:

Here we're seeing a 2.3% increase in average, and 9.5% in minimum.

Next, AVP:

Here it's a 5% increase in average, and overall frames increases by 5.1% as well.

If you've read this article it shows how the 8x lanes on a p55 mobo restrict a 5870 by about 4% because it's able to output that much more data than an 8x lane can hold. My OCed 5850s should be using up at least as much if not a bit more than stock 5870s, so this is a good way to gain back a bit of that loss.
Could be that there's still a CPU bottleneck, for example in Vantage going from a 3.68ghz OC with turbo (up to 4.2ghz) to a straight 4ghz no turbo OC, while I gained 2k in CPU score I lost 3k in GPU score. However, that's also DX10. Still, perhaps I'll find the time to rerun some of these tests with the turbo OC and see if it helps anything. (FYI, even AvP showed a slight performance decrease from the turbo OC to the 4ghz OC and it is DX11 - from 56.1 to 55.2)

EDIT: Unrelated to the PCIe frequency, but here's the Vantage shots:
175x20 with Turbo on:

200x20 straight 4ghz:

 

[Mugz]

On the earlier Core2 and P4 chips, upping the PCIE frequency used to allow a higher FSB overclock. I have not yet played around with the Core i-series yet, but I would not be surprised if that is still the case.

The reason it loses the network adapter and the sound card is because they are on the PCIE bus as well, and going so far out of tolerance makes them incapable of operation...

Nice work. I am surprised that it is only worth about 1-odd per cent improvement, though...

 

[Deleted member 217926]

Nicely done. That is the first time I have seen it done since the P4 days and very early PCI-E.

 

[Onus]

Nice, thorough post. I was under the impression that too high a PCIE frequency could actually damage cards though. Is this dated, or simply wrong?

 

[Mugz]

Upping the PCIE frequency has pretty much the same effect as upping the FSB, only on different components. All that happens when the frequency goes out of tolerance is the component whose tolerances have been exceeded stops responding.

 

[wolfram23]

Thanks for the replies

jtt283 - to be honest, I don't know. However, I can't really see why it should cause damage. Worst case scenario is that something doesn't work right at a higher frequency... but I guess time will tell. I just want at least, oh, 3-4 years out of my PC so I'll find out first hand if overclocking kills lol.

Mugz - I expected a slightly higher result as well. I'm going to play around with my GPU OC, it's possible that now I can OC it more... who knows. I'm also going to try going back to a Turbo OC, and see if I can push the base clock to at least 180, which I was unable to do before (at least without significantly more voltage + higher temps than I was comfortable with)


BTW, any comments on the Vantage GPU score dropping when going to a higher straight OC compared to the Turbo OC?

 

[wolfram23]

Planned future update: Rerun the FurMark, AvP, and Dirt 2 tests with the Turbo overclock (175bclk) to see if there's a FPS increase while at 130mhz PCIe.
Secondly, I will see if I can get a higher base clock stable OC. I should probably be able to do this over the weekend.

I'm a little surprised at the lack of interest in this, I kinda thought this was some new info lol - at least it's not something talked about a whole lot.

 

[4745454b]

Last I heard people would change the PCIe bus to help improve their "FSB" OC. Remember that some systems (at least back in the P4 days) would tie the various buses together in some way. By using ratios they could then drop X bus back to where it should be. Sometimes you'll hit a ratio that doesn't work well. By changing the frequency you'll enable a different ratio which might allow you to hit a higher FSB. I haven't seen anything more about this since the C2D series came out.

I'm actually rather impressed with this. Nice thinking outside the box. I'm not surprised at the lack of increase in speed. 16x is more then enough to send data to the card, this is simply proof that we don't need a bigger bus, even for the 5850.

 

[wolfram23]

^ Thanks

I got the idea when I read a review a while back that sort of mentioned it but didn't do any testing. My main objective was to be able to get the GPU usage higher and smoother as it sometimes drops to, say, 95% usage from 99/98%.

As a side note, on my GT 240 PhysX card smoothed out it's usage wonderfully when I increased the PCH voltage from 1.05 to 1.15V (tested with Fluid Mark).

 

[jsc]

I'm really not surprised that you see very little improvement. It just confirms that the rendering engine in the video card is the bottleneck, not the PCIe interface.

 

[OvrClkr]

I get maybe a frame or two @ 120Mhz, nothing earth shattering so I just leave it at 100Mhz and focus more on the GPU clocks.

 

[Andon48]

I have always wondered what the point of the PCI frequency was. I have never touched it, but I have hit a wall with my i3 and my 5750. I can't seem to go past 185 Bclk on the i3, or 970 core on the 5750. I wonder if upping the PCI frequency will help any? I'll give it a try and let you know what happens.

 

[Shadow703793]

I am going to bring this thread from the dead.

WTH was this buried???? I never saw this till today on another thread (ref: http://www.tomshardware.com/forum/262287-29-pcie-frequency ).

 

[wolfram23]

Haha... it's aliiiive! It's aliiiive!

But yeah, I've found a few real gems buried in this forum. Ah well!

 

[RJR]

Wolf, what happened with the rest of the testing?????

 

[wolfram23]

Umm.... I don't know!

Well, my PC wasn't booting the next day until I dropped down to 120 or 115mhz PCIe so I guess that kind of put a wrench in it.

To be honest, at this point I'm quite happy with where my PC is so any free time I get I'm mostly just playing some games, however, I would definitely consider doing some more testing if there's anything worth looking into. At this point the only thing left I feel is either trying to break 4ghz on 4 cores or else trying to push the turbo OC higher, but that would just be for fun.

 

[RJR]

OH, OK, I was curious to see if it ever led to anything good.

Was hoping like a 4.2 Ghz OC at 1.2v or something.

 

[wolfram23]

Oh that would have been nice! Maybe I'm crazy, but I do think a higher base clock allowed a slightly higher PCIe frequency, because I was running those tests at 200x20 but for normal use I'm at 177x20 with turbo... Hmm...

 

[swesen]

you have counted all the tests wrong. avp for example 6076/5776=1.05 which is 5% increase not 1.05%, 1=100%, 0.05=5%.

 

[wolfram23]

Best answer selected by Wolfram23.

 

[wolfram23]

swesen - yeah I think you're right. Funny that nobody else pointed that out.