QuadGT PCI-E

[mirek]

Hi, I would like to know what is the diffrence betweem PCI-E x16 and x4. Does it mean if I put crossover graphic cards, both will work x4? Will this have any impact, or this speed doesn't have any role?
Thanks!!!

[gilgamesh]

dont forget a system will only run as fast as its slowest part. SO yes in theory they would both run at 4x

gilgamesh

[mirek]

I am not really familiar wtih this issue. I suppose that x16 is 4 time faster than x4, but I would like to now would it have lot of impacts to games performance. Ii it better to have one x16 card, or 2 as crossover X4 cards (you said, competer works as fast as its slowerest part does).
Thanks!!!

[Lucio]

I think you have to bear in mind that whilst there's an impact on the theoretical limits, actual real world performace varies greatly.

As this article shows http://hardocp.com/article.html?art=...50aHVzaWFzdA== the expected boost never materialised because the software isn't designed to make use of both cards.

[Johnny Bravo]

I'd recommend you read this before making any decisions it'll explain what is going on with a x16/x4 setup and the real world hit that you're gonna take if you decide to run crossfire

http://www.firingsquad.com/hardware/...ce/default.asp

I think it's incorrect to say they will both run at x4. The secondary PCI-e bus is basically running off the SB so what is actually happening is that a full x16 path is wired to the NB meaning that your primary gpu is running at x16 and the secondary graphics card is running at x4 BUT it has to go via the southbridge to the norhtbridge as well. The general consensus is that 8x is fine but x4 is a bit of a bottleneck so the improvements you will see from a crossfire setup under these conditions will be slightly under par with that of a x16/x16 or x8/x8. However the real world performance hit will vary and depending on what you play may not be that far off the normal crossfire speed. It basically will boil down to driver implementation and how they assign the resources available, ie 75% of work goes to the primary gpu, 25% to the secondary instead of 50/50 etc

[Agent]

I think it's incorrect to say they will both run at x4. The secondary PCI-e bus is basically running off the SB so what is actually happening is that a full x16 path is wired to the NB meaning that your primary gpu is running at x16 and the secondary graphics card is running at x4 BUT it has to go via the southbridge to the norhtbridge as well.

Doesn't that purely depend on the implementation of the PCI-e slots? I wouldn't say that is universally true.
Even so, ATI has a composite engine on the cards them self (On the master card / slave setup), so this would take out the need for cross-bridge communication as it doesn't pass the data over the PCI-e bus (965 doesnt support peer-to-peer writing), surely?

It basically will boil down to driver implementation and how they assign the resources available, ie 75% of work goes to the primary gpu, 25% to the secondary instead of 50/50 etc

With the master / slave setup that is needed (thats the only way to do crossfire on 965) that simply isn't true.
The work load will still be shared fairly evenly - its the reason why ATi (opps...AMD) went with the dongle solution, specifily to avoid the situation you are saying
Heck, the article you have linked shows that with the 6.10 drivers there is a maximum of 1FPS between the dongle solution on the 965 chipset and the 975 which has the ability to peer-to-peer write on the PCI-e bus

[Johnny Bravo]

Doesn't that purely depend on the implementation of the PCI-e slots? I wouldn't say that is universally true.
Even so, ATI has a composite engine on the cards them self (On the master card / slave setup), so this would take out the need for cross-bridge communication as it doesn't pass the data over the PCI-e bus (965 doesnt support peer-to-peer writing), surely?

While it does depend on the implementation of the board it's usual if the motherboard manufacturer has chosen to put a mechanical x16 slot on the board then they expect that some users will use it for a PCI-E graphics card. Now there's the catch, yes you're right when you say the dongle is the alternative way ie either write via p2p or the dongle but from what I gather ati is moving towards removing the making it alot more like the SLi application. The 965 does not do peer to peer so Ati uses the Direct Media Interface (DMI) instead hence the whole NB ->SB link.

With the master / slave setup that is needed (thats the only way to do crossfire on 965) that simply isn't true.
The work load will still be shared fairly evenly - its the reason why ATi (opps...AMD) went with the dongle solution, specifily to avoid the situation you are saying
Heck, the article you have linked shows that with the 6.10 drivers there is a maximum of 1FPS between the dongle solution on the 965 chipset and the 975 which has the ability to peer-to-peer write on the PCI-e bus

Ok firstly I don't understand exactly your reasoning here, ati like nvidia use a number of load sharing modes to split the image load, and in accordance with the drivers it can share the load based on the electrical setup too.

Though I guess I should have stated that I'm not saying it will always run at 75/25 I was getting the idea that the load gets split between the card unevenly when it needs to be, therefore its not locked at 50/50 and by extension both cards are forced to work at the limitation of the x4 bus.

In regards to the article linked what it shows is that if the drivers are well written then the difference between the number of lanes available to the graphics cards doesnt mean all that much at the end of the day!

[BUFF]

I'm sure that Anandtech did 2 articles covering this (on 965) but I'm damned if I can find them now.
Anyway iirc by the 2nd one the drivers had improved & there wasn't much performance loss at all compared to e.g. a dual x8 config..