*Asus P8P67 (PCIe 16x/4x) vs. Asus P8P67 Deluxe (PCIe 8x/8x) - HD6950s in CF tested*

[RussianSensation]

Browsing the forums, I noticed that some posters still thought that 8x/8x neutered videocard performance compared to the full fledged 16x/16x, which of course Socket 1366 setup offers.

Just a quick reminder, the PCIe 2.0 8x/8x vs. PCIe 2.0 16x/16x was already shown to be pretty immaterial. I specifically like the analysis performed by TechPowerup found here, where they found the performance difference between x8 2.0 and x16 2.0 was about 1-2%. Alternatively, HardOCP even investigated this at higher resolutions (busydude, thanks for the link).

LegitReviews has just ran a comparison of two P67 boards, which included HD6950s in Crossfire. The findings were definitely not what I had expected.

Asus P8P67 = 16x/4x
Asus P8P67 Deluxe = 8x/8x

Benchmarks






Source.

How can this be? My 2 cents:

The new Intel P67 chipsets have proper full PCI Express 2.0 bandwidth, doubling the data throughput available compared to what was available on P55. This means PCIe 2.0 4x on P67 should be equal to PCIe 1.0 8x on P55 (feel free to correct me here). Perhaps, this is why we seeing such a small performance drop. The other possibility is that HD6950 is just not fast enough and/or those games aren't demanding enough at 1920x1080 to saturate the bandwidth.

In any case, PCIe 2.0 8x/8x on P67 should definitely be a non-issue for those planning to upgrade.

[NumericalMethods]

Hmm, I'm confused.

Firstly both P55 and P65 boards supply 16 PCI-e 2.0 lanes for graphics. So there shouldn't be any difference in graphics bandwidth. Intel chipsets have supported PCI-e 2.0 since P45 (or was it x48?). I agree with the statement that x4 PCI-e 2.0 should provide approximately the same bandwidth as x8 PCI-e 1.0

Secondly, I see then benchmarks at 1080p varying very little between the x16/x4 and x8/x8 configs, which is interesting. Even the x4 lane doesn't seem to ruin the crossfire performance.

Third, the ECS P55 figures are pretty bad compared to those of the P65. I would have assumed they should be much more similar. Then I looked up the board and found it has an NF200 bridge chip to provide more PCI-e lanes. I'm guessing that there could be a configuration issue with that test setup as the numbers seem off to me.

Finally, I 100% agree that PCI-e 2.0 8x/8x should be a non-issue for all but the pickyest of enthusiasts upgrading to P67. I personally would still be a bit worried about potential performance loss on a x16/x4 setup however.

Does any of that contribute something useful?

[=Wendy=]

I believe, and someone will correct me if i'm wrong, but although P45 and P55 supported PCIe 2.0, the bandwidth was uni-directional, but only now with P67, is PCIe 2.0 bandwidth bi-directional.

[Zap]

Originally Posted by NumericalMethods

Third, the ECS P55 figures are pretty bad compared to those of the P65. I would have assumed they should be much more similar. Then I looked up the board and found it has an NF200 bridge chip to provide more PCI-e lanes.

NF200 is not perfect. It adds a bit of latency, plus doesn't really increase the number of PCIe lanes - it really seems to just load balance them. For instance, yes, NF200 on a P55 chipset board will give each slot a "full x16" but of course the NF200 still communicates to P55 through x16 on its own. To use an analogy, if a typical house circuit supports 15A and you use two surge strips that each supported 15A, then each surge strip supports 15A but the whole circuit still only supports 15A.

Wait, if someone didn't understand the PCIe lanes, then they won't understand circuits.

Originally Posted by =Wendy=

I believe, and someone will correct me if i'm wrong, but although P45 and P55 supported PCIe 2.0, the bandwidth was uni-directional, but only now with P67, is PCIe 2.0 bandwidth bi-directional.

That's interesting. I've heard people mention that P67 has more bandwidth, but everyone seemed to be parroting everyone else, and nobody had an explanation as to why.

[coffeejunkee]

Thanks for those benchmarks.

I thought the Techpowerup test was good, at least for x16/x16-x8/x8 comparison.

But it didn't take into account the fact that on P55 the x4 slot is connected to the pch and not to the pci-e controller in the cpu itself. I expected this to be noticeably slower.

But looking at these numbers it doesn't really seem to matter. Now it would be interesting to see this tested on P55 with the lower pci-e 1.1 bandwidth.

[NumericalMethods]

Originally Posted by =Wendy=

I believe, and someone will correct me if i'm wrong, but although P45 and P55 supported PCIe 2.0, the bandwidth was uni-directional, but only now with P67, is PCIe 2.0 bandwidth bi-directional.

I didn't know that, good knowledge. Was the bandwidth in the opposite direction only PCI-e 1.0 spec if that was the case?

Originally Posted by Zap

NF200 is not perfect. It adds a bit of latency, plus doesn't really increase the number of PCIe lanes - it really seems to just load balance them. For instance, yes, NF200 on a P55 chipset board will give each slot a "full x16" but of course the NF200 still communicates to P55 through x16 on its own. To use an analogy, if a typical house circuit supports 15A and you use two surge strips that each supported 15A, then each surge strip supports 15A but the whole circuit still only supports 15A.

Sorry, I was aware of how the NF200 chip is electrically connected. I agree that in theory the increased bandwidth of 16x/ 16x card to card should improve bandwith and therefore performance (or not as 8x/ 8x has been shown to be perfectly adequate). I was more inferring that there might be a setup issue between the NF200 chip and crossfire since that figures are really low on the ECS board and I didn't think they should be. Sorry for the confusing post.

[=Wendy=]

The bandwidth for P55 and P67 is identical. (16GB/s - 2.5GT/s)
It is my understanding that in P55 you could either be sending a data packet from the PCIe to the CPU, or receiving a data packet from the CPU to PCIe, you couldn't send and receive a data packet at the same time.

In P67, it my understanding that you can send and receive a data packet at the same time, so in theory you double the throughput on P67.

Actually, it's not P67 alone, it has more to do with how Sandy Bridge works compared to how Lynnfield works.

[RussianSensation]

Originally Posted by =Wendy=

The bandwidth for P55 and P67 is identical.

There are actually two important characteristics associated with PCI Express: (1) data rates, and (2) transfer rates.

P67 chipset diagram.
The CPU has 16 lanes of PCIe 2.0 for a total data rate bandwidth of 16GB/sec.
PCIe 2.0 lane by definition has a transfer rate of 5.0GT/sec and runs at a speed of 5.0 Ghz.
The P67 chipset has 5GB/sec of data rate bandwidth x 1.
P67's chipset PCIe lanes operate at a full PCIexpress 2.0 transfer rate speed.

P55 chipset diagram.
The CPU has 16 lanes of PCIe 2.0 for a total data rate bandwidth of 16GB/sec (so 1156 and 1155 CPUs don't differ here).
However, on the P55 chipset, each PCIe lane only offer signaling/transfer rates up to 2.5GT/s (which means they probably operate at 2.5 Ghz only), which is the same speed as PCI Express 1.0 (gen one).
The P55 chipset has 0.5GB/sec data rate bandwidth x 1.
P55's chipset PCIe lanes operate at PCIexpress 1.0 transfer rate speeds.

Low to mid-range P55 boards would run dual graphics cards in 16x/4x such that the first PCI Express slot wqould use the lanes from the CPU and the second PCI Express slot would use the lanes from the P55 chipset.

P67 chipset offers 2x the transfer rate over P55. Therefore, what was PCIe x4 on P55 is essentially as fast as PCIe x8 on P55 from a transfer rate perspective. However, this is still called PCIe x4 on P67 but the transfer rate is twice as fast as it was on P55 under the same x4 lanes.

This may explain why P67 x4 is not bottlenecking the HD6950 but P55 x4 would.

[realbabilu]

Please check the bandwith with these:
For ati: http://developer.amd.com/Downloads/P...dTest_v0.2.zip
For nvidia [opencl card only] OpenCL Bandwidth Test
x86 http://developer.download.nvidia.com...dwidthTest.zip
x64
http://developer.download.nvidia.com...dwidthTest.zip
For all cards: http://www.kegetys.net/dl.php/D3Dbandwidth.zip

I believe ATI pciespeedtest is the best solution for measure, unfortunately not capable for measure other than ati.

[RussianSensation]

TechReport just did a quick review of P67 PCIe x1 vs. P55 PCIe x1.
http://techreport.com/articles.x/20241

Considering the Syba add-on card's Marvell controller was still holding back the 2 drives in RAID (as linked in the review), the PCI express performance can be safely assumed to have doubled!!

Thus, P67 chipset offers twice the PCI express bandwidth over P55 chipset for add-on cards like SATA 3.0, USB 3.0, PCIe videocards, etc.

[Castiel]

Originally Posted by RussianSensation

There are actually two important characteristics associated with PCI Express: (1) data rates, and (2) transfer rates.

P67 chipset diagram.
The CPU has 16 lanes of PCIe 2.0 for a total data rate bandwidth of 16GB/sec.
PCIe 2.0 lane by definition has a transfer rate of 5.0GT/sec and runs at a speed of 5.0 Ghz.
The P67 chipset has 5GB/sec of data rate bandwidth x 1.
P67's chipset PCIe lanes operate at a full PCIexpress 2.0 transfer rate speed.

P55 chipset diagram.
The CPU has 16 lanes of PCIe 2.0 for a total data rate bandwidth of 16GB/sec (so 1156 and 1155 CPUs don't differ here).
However, on the P55 chipset, each PCIe lane only offer signaling/transfer rates up to 2.5GT/s (which means they probably operate at 2.5 Ghz only), which is the same speed as PCI Express 1.0 (gen one).
The P55 chipset has 0.5GB/sec data rate bandwidth x 1.
P55's chipset PCIe lanes operate at PCIexpress 1.0 transfer rate speeds.

Low to mid-range P55 boards would run dual graphics cards in 16x/4x such that the first PCI Express slot wqould use the lanes from the CPU and the second PCI Express slot would use the lanes from the P55 chipset.

P67 chipset offers 2x the transfer rate over P55. Therefore, what was PCIe x4 on P55 is essentially as fast as PCIe x8 on P55 from a transfer rate perspective. However, this is still called PCIe x4 on P67 but the transfer rate is twice as fast as it was on P55 under the same x4 lanes.

This may explain why P67 x4 is not bottlenecking the HD6950 but P55 x4 would.

Russian,

I'm sure you've heard this before but you're one smart mofo

[RussianSensation]

Originally Posted by Castiel

Russian,

I'm sure you've heard this before but you're one smart mofo

Thanks for that Castiel. I am not that smart. I am simply using the information I gathered to help myself better understand what I don't know. Hopefully that will result in a discussion where we all can learn something new.

You settled on 4.6ghz on that 2600k? That's a nice overclock at 1.320V. Can the chip do 4.7ghz @ 1.35V? Are you using the stock cooler? What are your temperatures at load?

[=Wendy=]

PCIe2 lanes connected to the P67 have 5Gb/s (gigabits per second) not 5GB (gigabyte) per second per x1 lane, and there are 8 lanes, so that is 40Gb/s total bandwidth. However it must use DMI to communicate with the CPU. In the P67 diagram, DMI has 20Gb/s.
So something doesn't add up.

[Castiel]

Originally Posted by RussianSensation

You settled on 4.6ghz on that 2600k? That's a nice overclock at 1.320V. Can the chip do 4.7ghz @ 1.35V? Are you using the stock cooler? What are your temperatures at load?

I just haven't had time to fool around with it this week. I can boot into windows at 4.8ghz with 1.35v but it needs a touch more for stability. I'm running an H70. Idle's are 26-28C, Full load's in the high 50's

[jjdeltor]

Hello, I was reading this thread and had a question about PCI-E lanes on the P67 build I am doing (Asus board, either the pro or the deluxe). How does the use of both a discreet GPU and a dedicated PhysX affect lane allocation? I have a GTX480 and my brother gave me an old 9600 GT to use should I want to. Would I be able to keep the GPU (GTX480) still running at 16x and run the dedicated PhysX card (9600 GT) to my benefit or would there be some sort of bottleneck? I am new to having a dedicated PhysX card and was just curious as to how to set this up to my advantage given the motherboard I am going to be using. Thanks for any help you can send my way.

- J

[Castiel]

Originally Posted by jjdeltor

Hello, I was reading this thread and had a question about PCI-E lanes on the P67 build I am doing (Asus board, either the pro or the deluxe). How does the use of both a discreet GPU and a dedicated PhysX affect lane allocation? I have a GTX480 and my brother gave me an old 9600 GT to use should I want to. Would I be able to keep the GPU (GTX480) still running at 16x and run the dedicated PhysX card (9600 GT) to my benefit or would there be some sort of bottleneck? I am new to having a dedicated PhysX card and was just curious as to how to set this up to my advantage given the motherboard I am going to be using. Thanks for any help you can send my way.

- J

Spend the extra money if considering the deluxe

http://www.newegg.com/Product/Produc...-694-_-Product

[jjdeltor]

Is this a suggestion on just a price point (I.E. the small price difference between the Deluxe and the Work Station board), or this a suggestion concerning my GPU/PhysX and PCI-E lanes question?

- J

[Castiel]

Originally Posted by jjdeltor

Is this a suggestion on just a price point (I.E. the small price difference between the Deluxe and the Work Station board), or this a suggestion concerning my GPU/PhysX and PCI-E lanes question?

- J

Both. Thats the cheapest board with an NF200 chip allowing for X16 X16 X8 3 way Xfire or SLI

[VirtualLarry]

My understanding was that the PCI-E lanes coming off of the CPU, on both 1155 and 1156, are PCI-E 2.0, either x8/x8 or x16.

The PCI-E lanes coming off of the chipset, were only PCI-E 1.0 speed on 1156, but are PCI-E 2.0 speed on 1155.

[divertiti]

sorry to revive this thread, but I'm really interested in this topic. There's a new review on this topic done recently, and the results quite contradict those from Legit Review even though they were using pretty much the same hardware...any idea why?

[RussianSensation]

Originally Posted by divertiti

sorry to revive this thread, but I'm really interested in this topic. There's a new review on this topic done recently, and the results quite contradict those from Legit Review even though they were using pretty much the same hardware...any idea why?

Actually if you look at that review, outside of F1 2010, the performance difference for most games was only about 10% (obviously not optimal), but clearly not prohibitive enough to discount the option of getting a 2nd HD6950 for a 16/4x setup (CF generally will add 70-90% scaling under optimal circumastances). So even after a 10% adjustments, you would still get a 60-80% scaling out of a 4x slot. Not bad for a budget board. Having said that, obviously if a person is spending $500+ on a videocard setup for their initial build, then of course getting at least an 8/8x motherboard is the way to go.

[divertiti]

Originally Posted by RussianSensation

Actually if you look at that review, outside of F1 2010, the performance difference for most games was only about 10% (obviously not optimal), but clearly not prohibitive enough to discount the option of getting a 2nd HD6950 for a 16/4x setup (CF generally will add 70-90% scaling under optimal circumastances). So even after a 10% adjustments, you would still get a 60-80% scaling out of a 4x slot. Not bad for a budget board. Having said that, obviously if a person is spending $500+ on a videocard setup for their initial build, then of course getting at least an 8/8x motherboard is the way to go.

I realize F1 is kind of an outlier, however even considering the other games, LR showed almost no difference at all, that's a pretty big stretch from TH showing 8-10% difference on the same games and same hardware...

[Fanatical Meat]

The more things change, the more they stay the same. This reminds me of the agp 2/4/x8 debate, no real difference between 4 & 8 and a small hit on 2.

[RussianSensation]

Originally Posted by divertiti

LR showed almost no difference at all, that's a pretty big stretch from TH showing 8-10% difference on the same games and same hardware...

Look more closely. The 8-10% average is only because F1 2010 is skewing the results. The performance difference is closer to 4-7%.

AvP with 4AA 1920x1080

Legit Reviews:
PCIe 4x = 71.3 fps
PCIe 8x = 75.1 fps (+5.3%)

Tom's Hardware:
4x = 71.9 fps
8x = 76.8 fps (+6.8%)

The difference between these 2 is only 1.5% (well within the margin of error)

STALKER: cop

Legit Reviews:
4x = 121 fps
8x = 124.5 fps (+2.9%)

Tom's Hardware:
4x = 117.3 fps
8x = 121.5 fps (+3.6%)

So a 0.7% difference....

In Crysis, Tom's hardware found 0.2 fps difference.

It depends on the game. The point is for the majority of games 16x/4x CF on the newer P67 chipset is not much slower than 8x/8x CF. Even with a 10% deficit, it's not enough to prohibit someone from buying a 2nd videocard which will generally boost performance about 70-80%.

[grandpatzer]

I'm considering P8P67 Deluxe B3 to run 6950 in tri-fire (x8 x8 x4) but it seems as this might not be a good idea on 5760*1080 resolution.

I'm currently on m4a79t AM3 system and can can do dual 16x/16x, triple 16x/8x/8x or quad 8x/8x/8x/8x.

Seems best to just hold on to the phenom ~4ghz and tri/quad 6950?