Sandy Bridge-E benchmarks on Techpowerup.com

[Edrick]

I guess I wasn't very clear, so I'm going to try this again. There is only a single set of 16 lanes coming from SNB. NF200 changes how you can allocate bandwidth, but you will never get more than 16 lanes worth of bandwidth (8GB/sec each way) out of SNB. NF200 can help by keeping the 3rd PEG slot from being gimped at x4, but it can't come up with more bandwidth from SNB.

Exactly.

 

[MrX8503]

Why do gamers expect people who build workstations to only be gaming? This I never understood. PCs aren't just toys, people use them for actual work too.

 

[antisocialmunky]

Why do gamers expect people who build workstations to only be gaming? This I never understood. PCs aren't just toys, people use them for actual work too.

You can't run 1 game faster but you can run 3 copies of one game

 

[RussianSensation]

Russian there are 16 lanes no matter what with SB. The NF200 just tricks it out to allow you that third card.

I will dig up the review I read explaining it all. Think it was over at [H].

I guess I wasn't very clear, so I'm going to try this again. There is only a single set of 16 lanes coming from SNB. NF200 changes how you can allocate bandwidth, but you will never get more than 16 lanes worth of bandwidth (8GB/sec each way) out of SNB.

NF200 can help by keeping the 3rd PEG slot from being gimped at x4, but it can't come up with more bandwidth from SNB.

Exactly.

NF200 chip acts as a multiplexer. P67 natively has 16 lanes of 2.5GT/s PCI-E 2.0. The NF200 takes 16 lanes and turns them into 32.

Did you guys read any of the reviews I linked? One of them is by HardOCP.

If the P67/Z68 platform can only do 16x lanes in total (8x/8x), then it would be impossible to run 3 GPUs at 8x/16x/16x. But this is not the case.

When you have 3 graphics cards, "The first slot will still be working in x8 mode but the other two, in full-speed x16 mode." ~ Xbitlabs

* Same thing stated by HardOCP, Guru3D, etc.

Also, just look at Hard OCPs 2500k/2600k Tri-SLI benchmarks. The 3 GPUs are not bottlenecked. The argument that P67/Z68 chipset + NF200 chip bottlenecks 3 GPUs is not correct.

Technicalities aside, the NF200 chipset fixes the PCIe issue that plagues P55/P67 and Z68 chipsets for Triple-GPU performance:

You can choose the X79 platform because it has 8 Dimms, supports SB-E, etc, but the performance difference between 3x GTX580s on P67/Z68 + NF200 vs. 3x GTX580s on X79 is going to be ~ 2-3%.

The P67 chipset also doubles the bandwidth throughput vs. the P55, alleviating the PCIe bottleneck even further from the P55 platform.

P55 + NF200 had a 2% difference in Tom's Review on 3x HD5870
P67/Z68 + NF200 platform at HardOCP with 3x 580s smoked an overclocked i7 920 on X58
Same thing against a Bulldozer.

If the PCIe bandwidth was a real issue, then a 2500k/2600k would have no way of beating the other platforms with more native PCIe lanes.

 

[Edrick]

Also, just look at Hard OCPs 2500k/2600k Tri-SLI benchmarks. The 3 GPUs are not bottlenecked. The argument that P67/Z68 chipset + NF200 chip bottlenecks 3 GPUs is not correct. Technicalities aside, the NF200 chipset fixes the PCIe issue that plagues P55/P67 and Z68 chipsets for Triple-GPU performance:

Yes, it does help the issue of only having 16 PCIe lanes on P67/Z68 systems.

Yes, tri-SLI on that system only suffers a very small penalty as opposed to X58/X79 (~2&#37.

No one is really argueing that. But the fact is that the NF200 chip does not magically add more PCIe lanes to the CPU (only 16). It does add PCIe lanes to the system, which then in turn has to merge them into the 16 lanes on the CPU. And once GPUs get faster in the next generation, and start supporting PCIe 3.0, I bet users will see that 2% difference start to become larger. And that is when X79 will start to shine.

 

[RussianSensation]

No one is really argueing that. But the fact is that the NF200 chip does not magically add more PCIe lanes to the CPU (only 16). It does add PCIe lanes to the system, which then in turn has to merge them into the 16 lanes on the CPU. And once GPUs get faster in the next generation, and start supporting PCIe 3.0, I bet users will see that 2% difference start to become larger. And that is when X79 will start to shine.

I never said NF200 chip adds more lanes to the CPU, it's about the overall PCIe bandwidth for the entire system, which does increase.

There is only a 2% difference between PCIe 2.0 x8 and x16. So PCIe 3.0 isn't going to magically add more performance. That would only be true if there was already a large performance delta between 2.0 x8 and 2.0 x16. So clearly, even 2.0 x8 with half the bandwidth isn't a bottleneck. How is doubling the bandwidth of 2.0 x16 going to suddenly give us more performance?

The PCIe myth keeps marching on. It was around since PCIe first arrived on the scene and rumors were circulating of PCIe 2.0 x16 supplanting the original version shortly. People started to get worried that PCIe 1.0 x16 wasn't going to be fast enough for next generation of GPUs. I see that the marketing guys have succeeded. PCIe 1.0 x16 still doesn't bottleneck a modern GPU and neither does PCIe 2.0 x8.

Maybe if the performance difference grows from 2-3% to 15-20%, then I would start to get worried.

The reality is that the higher CPU overclocking of 2600k/2700k is likely to be a larger factor in performance than any of the advantages of X79 platform. Reading all the reviews around, 3930/3960 chips struggle to overclock much past 4.7-4.8ghz.

So really: LGA 2011 CPUs don't offer any more overclocking headroom, games don't take advantage of more than 4 cores and PCIe bandwidth is not a current bottleneck, not even for 3 GPUs.

The main purpose of X79 platform is for workstation applications. It offers no advantages to gamers or overclockers. And the fact that 8 Dimm boards cost $300+ and the cheapest 6-core CPU is $550+ isn't doing it any favors.

X79 is actually worse than X58. It arrived 1 year late after 1155, not 1 year earlier of 1156 like the X58 did. CPUs overclock worse, and yet X58 CPUs overclocked better due to higher Vtt tolerance and the i7 920/930s ability to clock to 4.0ghz with minor CPU voltages, while most i7 860 struggled to get there without a huge voltage increase. I've seen plenty of i7 920/930s running 4.2-4.4ghz but almost no i7 860s past 4.2ghz on reasonable Vtt/CPU voltage.

If Intel wants its ultra-high end platform to be taken seriously it has to:

1) Release its high-end platform not much later than its mainstream platform (and preferably earlier to make those enthusiasts feel like they have the best platform for months before anyone else)

2) Have the most features and not lack on features when the boards sell for $300+

 

[Edrick]

The PCIe myth keeps marching on. It was around since PCIe first arrived on the scene and rumors were circulating of PCIe 2.0 x16 supplanting the original version shortly. People started to get worried that PCIe 1.0 x16 wasn't going to be fast enough for next generation of GPUs. I see that the marketing guys have succeeded. PCIe 1.0 x16 still doesn't bottleneck a modern GPU and neither does PCIe 2.0 x8.

I am certain that a GTX590 would bottleneck on a PCIe 1.0 x16 slot. Even if that is less than 8%, it still would.

Also, PCIe slots are used for more than just GPUs. Some of the newest PCIe SSDs will love this, and once they move to PCIe 3.0, then we will see some blazing fast SSDs. Not to mention RAID controllers and such. Sure most 'gamers' dont need that, but it does add value. (Tri-SLI with a PCIe SSD or 2 raided?)

 

[grkM3]

take 3 gtx590s,overclock them to 1.5 ghz(to sim next gen cards)and run them on a p67 and then on a x79,Ill bet there is more than a 5% gain.

 

[ed29a]

And once GPUs get faster in the next generation, and start supporting PCIe 3.0, I bet users will see that 2% difference start to become larger. And that is when X79 will start to shine.

Doubt. x16 + x16 vs x8 + x8 are trading blows in crossfire 6990 + 6970. In high resolutions, the difference is something minuscule like 1 fps (around 0.5% to 1.3% depending the test).

 

[Edrick]

take 3 gtx590s,overclock them to 1.5 ghz(to sim next gen cards)and run them on a p67 and then on a x79,Ill bet there is more than a 5% gain.

I am not certain you can TRI SLI the 590s since they are already internally SLI'd

 

[ed29a]

take 3 gtx590s,overclock them to 1.5 ghz(to sim next gen cards)and run them on a p67 and then on a x79,Ill bet there is more than a 5% gain.

I see what you did there ...

Why stop at video cards? Why not take a 2700K and overclock it to 10ghz too to 'simulate' next generation CPUs too! And then, take some duct tape and tape a few more PCIx slots to the motherboard too to simulate next generation motherboards!

 

[grkM3]

I see what you did there ...

Why stop at video cards? Why not take a 2700K and overclock it to 10ghz too to 'simulate' next generation CPUs too! And then, take some duct tape and tape a few more PCIx slots to the motherboard too to simulate next generation motherboards!

what are you talking about?next gen cards will be 20-30% faster than what we have now.

You can buy a gtx 590 now,Im sorry I didnt know you cant run 3 of them.People are clocking these to 1ghz while benching and 1.5 is right where the next gen cards will be at.

a 2700k cpu is the same freaking cpu as a sandy E,this was said to show the actual limits of the 16 lanes vs 40.You cant seem to understand the meaning of what I said and Im not even going to waste my time explaining it clearer.

you can mickey mouse the chipset as much as you want but the cpu can only handle 16 lanes

 

[ed29a]

what are you talking about?next gen cards will be 20-30% faster than what we have now.

You have benchmarked them?

1.5 is right where the next gen cards will be at.

[Citation needed]

 

[grkM3]

You have benchmarked them?


[Citation needed]

yes!!!







you still dont get it.

ps look up the rumors of the next gen 7970 card,its supposed to match the performance or come real close to the dual core 6990 and its a single chip pcie 3.0 card.

now use your imagination on why I said clock the gtx 590s to 1.5ghz to stim next gen cards.

 

[ed29a]

you still dont get it.

So you are saying that the 610ish Mhz GPU of the GTX 590 will reach 1.5 Ghz in the next generation? Does this involve a lot of prayers, voodo or illicit substances?

 

[grkM3]

So you are saying that the 610ish Mhz GPU of the GTX 590 will reach 1.5 Ghz in the next generation? Does this involve a lot of prayers, voodo or illicit substances?

NOOOOOO forget it,the next gen cards are going to be faster than what we have now and it would take a mssive overclock to simulate the next gen cards.

damn are you retarded? serioulsy?

a 7990 card is suppposed to match a dual gpu card in performance and thats a single gpu,take 2 next gen gpus and throw them on a card and then sli/crossfire them.

it would take 2 gtx 590 overclocked to the sky to keep up......DO YOU NOW GET IT ????

now that you got a better meaning go back and re read my first post about the bottle necking the 16 lanes of a 1155 sandy.


Do not post these kinds of over-the-top negative comments...it is inflammatory and insulting.
Re:

damn are you retarded? serioulsy?

Also, profanity is not allowed in the technical forums.

Please take some time to familiarize yourself with the following portions of the AnandTech Forum Guidelines:

1) No trolling, flaming or personally attacking members. Deftly attacking ideas and backing up arguments with facts is acceptable and encouraged. Attacking other members personally and purposefully causing trouble with no motive other than to upset the crowd is not allowed.

We want to give all our members as much freedom as possible while maintaining an environment that encourages productive discussion. It is our desire to encourage our members to share their knowledge and experiences in order to benefit the rest of the community, while also providing a place for people to come and just hang out.

We also intend to encourage respect and responsibility among members in order to maintain order and civility. Our social forums will have a relaxed atmosphere, but other forums will be expected to remain on-topic and posts should be helpful, relevant and professional.

We ask for respect and common decency towards your fellow forum members.

Administrator Idontcare

 

[Nemesis 1]

Here is a review using MSI M/B This bios needs work but its still better than anything out as of now,

http://www.guru3d.com/news/core-i7-3960x-processor--msi-x79agd65-review/

 

[Nemesis 1]

Isn't the whole ideal of a ring bus (multi-drop bus) to avoid the extra connects that would raise latency? It's more of an arbitration strategy.

The extra latency would more likely be attributed to the larger size of the cache.

Read AT article on Intels ringbus its well explained there.

 

[nitromullet]

P67/Z68 + NF200 platform at HardOCP with 3x 580s smoked an overclocked i7 920 on X58
Same thing against a Bulldozer.

That's pretty astounding how much an overclocked 2600K does for three way SLI. I'd really like to see HardOCP re-run those tests with their 3960X/X79 rig to see what (if any) benefit there is to having more native PCIe lanes.

 

[rgallant]

NF200 chip acts as a multiplexer. P67 natively has 16 lanes of 2.5GT/s PCI-E 2.0. The NF200 takes 16 lanes and turns them into 32.

Did you guys read any of the reviews I linked? One of them is by HardOCP.

If the P67/Z68 platform can only do 16x lanes in total (8x/8x), then it would be impossible to run 3 GPUs at 8x/16x/16x. But this is not the case.

When you have 3 graphics cards, "The first slot will still be working in x8 mode but the other two, in full-speed x16 mode." ~ Xbitlabs

* Same thing stated by HardOCP, Guru3D, etc.

Also, just look at Hard OCPs 2500k/2600k Tri-SLI benchmarks. The 3 GPUs are not bottlenecked. The argument that P67/Z68 chipset + NF200 chip bottlenecks 3 GPUs is not correct.

Technicalities aside, the NF200 chipset fixes the PCIe issue that plagues P55/P67 and Z68 chipsets for Triple-GPU performance:

You can choose the X79 platform because it has 8 Dimms, supports SB-E, etc, but the performance difference between 3x GTX580s on P67/Z68 + NF200 vs. 3x GTX580s on X79 is going to be ~ 2-3%.

The P67 chipset also doubles the bandwidth throughput vs. the P55, alleviating the PCIe bottleneck even further from the P55 platform.

P55 + NF200 had a 2% difference in Tom's Review on 3x HD5870
P67/Z68 + NF200 platform at HardOCP with 3x 580s smoked an overclocked i7 920 on X58
Same thing against a Bulldozer.

If the PCIe bandwidth was a real issue, then a 2500k/2600k would have no way of beating the other platforms with more native PCIe lanes.

""Finally, another shader heavy game really came to life with the new 4.8GHz system. GeForce GTX 580 3-Way SLI performed 16.7% faster with the new system versus the previous system. This is also astonishing given that we know how shader demanding this game is. Those three GTX 580's just can't flex their muscle without a fast CPU.""
just saying
-ha 4.8 vs 3.6 , 33% faster sb and only 16.7% faster in that bench.
-proves nothing unless they ran both at 4.2 the avg. oc on a 920

 

[nitromullet]

proves nothing unless they ran both at 4.2 the avg. oc on a 920

It actually does prove RussianSensations's point that P67+NF200 is not PCIe bandwidth limited compared to X58, and by extension presumably X79.

Evening out the cpu between the two would be helpful though... that's why I'd like to see the same comparison with the addition of a 3960X/X79 at the same clocks thrown into the mix.

 

[rgallant]

"Evening out the cpu between the two would be helpful though..."

-running 3 6970's could have helpful also
-I just never liked that bench , I don't care if it made the 920 look weak ,it just looks rigged every time I see it.
-also to run 3 580sli on a x58 might need some bios voltage tweaks .not plug and play for evga boards anyways.
-I'll have a look at ib when it comes out myself.

 

[Grooveriding]

The problem with that [H] review has always been that he is gimping the i7 and giving the SB such a huge overclock. 3.6 i7 vs the 4.8 SB ?

For Tri SLI/CF you need a big CPU. Why is he not using an i7 at 4.0-4.2 ? There is a breakpoint with your CPU where you will overcome the CPU bottleneck and open your cards up. Another 400-600MhZ may have done that for him.

Also as far as I know the MSI Eclipse X58 he used for the i7 runs at x16/x16/x4. x4 is a bottleneck for a high end card.

As far as the NF200 on the Z68 board. It does allow further PCIE lanes to open up before the CPU to get that third card. But at the end of the day, all those three cards still only have the 16x lanes to communicate with the CPU over. It cannot add more lanes into the CPU, only add additional lanes before the CPU to add the third card.

I am fairly confident that when we finally see a benchmark roll round of TRI-SLI on Z68 vs TRI-SLI on X79 with equally clocked SB chips. The X79 will be the faster platform.

 

[Idontcare]

As far as the NF200 on the Z68 board. It does allow further PCIE lanes to open up before the CPU to get that third card. But at the end of the day, all those three cards still only have the 16x lanes to communicate with the CPU over. It cannot add more lanes into the CPU, only add additional lanes before the CPU to add the third card.

This. Its a technical distinction but it merits being crystal clear about and not muddying the waters over whether or not it is needed, warranted, etc.

The extra PCIe lanes are in-effect added to the motherboard, not added to the CPU.

The GPU's in SLI/Xfire configuration do take advantage of the extra bandwidth native on the mobo in terms of communicating with each other.

The benefit is not zero. Whether or not the benefit is >>0 or simply >0 is a whole other discussion though.

But the bandwidth into and out of the CPU is not increasing just because an ASUS engineer bolted an NF200 onto my MIVE-Z. The bandwidth between the cards has, but not the bandwidth to the CPU.

 

[RussianSensation]

There is another problem with the LGA2011 CPUs - they have much lower temperature thresholds. The 3930 starts to throttle at just 86*C. If we are only talking about gaining the most optimal performance in games, then I have a feeling IVB will give us 5.3-5.5ghz overclocks on 28nm while most LGA2011 chips will still be stuck at 4.6-4.8ghz. So again, I feel that if an enthusiast is paying such a premium for the LGA2011 platform, it should undoubtedly be superior, at least for a foreseeable future.

It seems to me in just a matter of 5-6 months, it'll once again become 2nd best. With IVB-E arrive after IVB, I feel that Intel is not prioritizing the enthusiast market enough to warrant such lengthy delays for its most expensive platform.