PCI-express.

[ex_forum_user_3]

Hi there,

Because of too many questions about PCI-express, I will try to explain this very short and quickly.
Before asking questions, know that PCI-e can use any number of lanes, this is expressed as 1x, 2x, 4x, 8x, etc....any number you like up to the max of your bus.
This number has NOTHING to do with speed, forget the speed factor.

What this number means is very simple:

1x = 1 lane used
8x = 8 lanes used
16x = 16 lanes used

All lanes are equal as fast, but if you use more: the more bandwidth you have in total.
Another issue is PCI-e 1.1 versus PCI-e 2.0

PCI-e 1.1 = 250MB/s
PCI-e 2.0 = 500MB/s
PCI-e 3.0 = ~900MB/s (expected speed per lane)

So it's simple, just multiply the number of lanes you use with the speed per lane, that's your total speed.
However, your PCI-e bus will operate as max speed to the lowest factor, meaning put a 1x card in a 16x lane slot, it will do 1x.
Also, if the card is PCI-e 1.1, the bus will operate at PCI-e 1.1 speed, and not above.
Ergo, the minimal speed of a device or bus will determine it's max speed!

Also, the bus is designed (on purpose?) to work at any time, so if you have e.g. a 16x videocard, but it seems to operate at 8x AND you are sure everything is capable of doing 16x.
Then the card isn't making contact with all lanes and does a fall-back to the number of lanes that DO WORK.
Simply reseat the card and try again, also check SLI switch cards, as they could force a slot into 8x mode instead of 16x mode.

Anyway, to be short, the common dominator sets the speed, not any of the maxes.
And a badly seated card will work, but at lower x rating, make 200% sure it's properly fitted, as the x selection IS HARDWARE RELATED!

 

[MexicanSnake]

Interesting, I would like to get a PCIe version 3 but they are not avalible yet, they should have a superb performance.

 

[ex_forum_user_3]

Why? You can't even max out PCIe 1.0 :lol_anim:
Current stuff isn't fast enough.

 

[DaLoona]

Why? You can't even max out PCIe 1.0 :lol_anim:
Current stuff isn't fast enough.

Eeeeh I beg the differ, here's my opinon:

Been running 2 PCIe 2.0 cards, see sig, on 2x8 lanes PCIe 1.x, and its filled to the max and not using all the graphics power I have onboard.
Compared it to a mates pc, with 2 9800GTX+ on a PCIe 2.0 board, P7N Platinum, and he gets a big chunk o fps more on the same games, with similar clocks on the cards.
Im runnning cpu on 3.44GHz he was on 3.0GHz with an E8400, my mem runs 860MHz his 800MHz, same kit, so his setup is a little slower, I might add.
His board being almost identical, except for the fact he has PCIe 2.0 and can use Yorkfield cpu's and I dont.
And yes, all works fine and is properly installed, different testing points to the same conclusion, its full.

And Ill bet if you have a board, with 2x16 lanes PCIe 1.x, youll get those stuffed too, if you use 2xGTX275/GTX285/295, or 2x5850/5870/5970.
PCIe 2.0 should be fine, although with 2x 5970, it will be running at its limits too, when run @2x8 lanes.
Ill see if I can find that review somewhere, I think it was on Tom's hardware, were it said that modern cards are already cut short by PCIe 2.0 @x8 lanes.

So PCIe 1.x is done for vga cards, especially if you run, like me, 2x8 lanes PCIe1.x.
Its nice for budgetcards, but 2x250GTS @x8 will be hampered already, so yes you can out max current PCIe 1.x standards.

I dont care however, because for folding its perfect  .



Edit:

http://www.tomshardware.com/reviews/pci-express-2.0,1915.html

Not the one I meant, but it gives an idea 

PCie 1.x @x16 = PCIe 2.0 @x8.
PCIe 1.x @x8 = PCIe 2.0 @x4
As you can see, its still enough for older cards like the 9800GX2, but that already fills it up completely, its playable with those cards, very good, but that isnt the issue here.
Its about maxing it out, well it can as you can see, with ease.
And if you run current demanding games with current vga cards, lets face it the 9800GX2 is no springchicken anymore and it runs out of bandwidth already on PCIe 1.x @x16, you'll deffo need PCIe 2.0.
Furthermore, the fact that the 9800GX2 does run out of bandwith on PCIe 1.x @x16, supports my story about my cards, since a 9800GX2 is nothing more then 2 9800GTX on 1 pcb, in SLi.

So again, PCIe 1.x is done, finito, for the newer VGA cards, it just isnt gonna be enough anymore.

 

[ex_forum_user_3]

http://www.tomshardware.com/reviews/pci-express-2.0,1915-10.html

Where exactly is it maxed?
I see a little difference, but that is measuring error, I do not consider that being maxed, 1x is maxed, so much is proven....but not the rest.
Don't forget, less lanes is more signalling per lane, but the data isn't maxed.
You just proven that PCI 1.x 8x isn't even maxed, let alone 16 times...have serious doubts about 4x.
Sorry, you have to do better, give me a test setup with 8 or 16x PCI-e 1.x that's slower then PCI-e 16x, 8x or 4x.
Why didn't Tome take different boards?

 

[DaLoona]

Here's a quote of Anandtech from a piece about Intel being sued:

Lynnfield only offers single PCIe x16 connection coming from the CPU, which wouldn?t be enough to fully feed 2 high-end GPUs.

From:
http://anandtech.com/cpuchipsets/showdoc.aspx?i=3690

That cpu has 16 PCIe lanes and as you can see they cant deliver enough.
These guys really know what they are talking about, so I rest my case, PCIe 1.x is done for the newer graphicscards.

Interesting piece of info btw, that link.

 

[ex_forum_user_3]

Sorry but that article is about a screwed chipset-connection made by Intel.
It has nothing to do with the fact that PCIe 1.x is faster enough for any todays GPU.
Your CPU/memory simply isn't fast enough to do it.

Also that article isn't about modern systems, in fact it's about very old stuff.
There is nothing interesting about it.

Especially because you quote the wrong part:

Intel ?created several interoperability problems? with discrete CPUs, specifically to attack GPGPU functionality. We?re actually not sure what this means, it may be a complaint based on the fact that Lynnfield only offers single PCIe x16 connection coming from the CPU, which wouldn?t be enough to fully feed 2 high-end GPUs.

It means that everything the CPU has to do needs to pass the simple 16x CPU <-> chipset but yes you will overload that one.
But modern CPU's like the I7 and AMD K-series have a hypertransport-bus (Intel gave it a different name, but who cares), that bus is way faster then PCI-e and it's only turned into PCI-e at the chipset.
However it doesn't have a CPU <-> chipset bottleneck like shown in the paper, nor does that bus need to feed the memory to the CPU and back.

Sorry but your are totally off track with that article.

I still say PCIe 1.x is fast enough to handle any modern card today even in 2xSLI of 2xCrossfire as long as you don't drop below 8-lanes.

 

[Aaron]

Also that article isn't about modern systems, in fact it's about very old stuff.
There is nothing interesting about it.

As far as I skimmed it was more about Intel's anti-competitive behaviour rather than PCI-E bandwidth.

I still say PCIe 1.x is fast enough to handle any modern card today even in 2xSLI of 2xCrossfire as long as you don't drop below 8-lanes.

I agree.  I'd call for 1.0 x16 minimum during multiple card setups, though.  Not sure for ATI's 5800 series and NVIDIA's coming Fermi in CF/SLI, but definitely with the older card generations.  If 16 lanes is being used; you can't really bottleneck a VGA card.

At the higher resolutions the card's performance does lower with the reduction from x16 to x8, but this still isn't a direct comparison of PCI-E 1.x to 2.0.


DaLoona, I can understand your confusions and belief.  However, there's just no solid evidence that concludes that 1.0 doesn't have enough bandwidth.

If different boards were used with the same CPU and memory, only difference was chipset and PCI-E revision; then we would be in business.

 

[ex_forum_user_3]

Look at the test-setup of that machine at Tom's.
The CPU<->Chipset bus is the problem there, as soon as you need a lot of data to pass between CPU and memory the GPU will suffer.
And that is exactly what happens there with FlightSim.

The bottleneck in your sample is NOT PCI-e, it's the CPU<->Chipsetbus that get's overloaded.
Try same test with an I7 or a Phenom system it will not show a PCI-e overload, just like the other tests show as well.

This is exactly why I don't read Tom anymore, he's wrong a lot of times and doesn't even know it.
Tom spread for years that 3D Mark was the ultimate test, he was wrong.
Tom did powersupply tests, but he did them wrong.
Tom didn't discover nVidia cheating on GPU tests, other people did.
Tom didn't discover Intel cheating with Sysmark (based on the Intel address BTW) other people did.

Tomshardware is a brain-dead website like most other review sites.