Quad core vs Dual core CPU for xfire/SLI

[Trizmatic]

In light if the guru3D i7 review revealing the huge gains in xfire and SLI performance my question arises....what about quad cores? They only tested a dual core to compare to the i7. Has anyone seen a dual core vs quad core comparison to see if it makes a difference for SLI and xfire? I haven't come across anything yet..

 

[Bjorn_Of_Iceland]

http://www.guru3d.com/article/core-i7-multigpu-sli-crossfire-game-performance-review/19

 

[lemonadesoda]

Nice link

 

[3870x2]

looks like processor scaling really is keeping sli and cfx down. Also, GTX280s choke like a cheap whore on anything above 1920x1200...

 

[Bjorn_Of_Iceland]

From other graphs that I've seen, looks like multi GPU users (current GPUs at least) are the ones who would greatly benefit from an i7 platform.

single GPU on an i7 is another thing though...

 

[lemonadesoda]

You know, reviewing the link in more detail... especically the OTHER pages earlier in the article, shows some issues in the tests:

1./ 2x SLI GTX260 is FASTER than some charts with 3x SLI GTX280

I suspect there was a problem with 3x SLI setup on the nForce chipset that X58 solves.

So yes, performance is MUCH BETTER on x58 with i7. BUT something isnt working out with the comparison against 3x SLI.

 

[Deleted member 3]

Is it the CPU or huge memory bandwidth that makes the difference?

 

[lemonadesoda]

If it *IS* really only down to memory bandwidth, then that's really interesting.

• AMD made the right decision to adopt onboard memory controllers early on.
• AMD is in big trouble, because their "cores" must be rubbish, but the difference has been hidden due to a real win on the memory bandwidth side. But that is now over. They have some serious work to do.
• BUT, IIRC, cache is still quicker than DDR3 latency and bandwidths, therefore, cache is king, and we should see bigger FPS scaling on different cache sizes on Core 2, but we dont.

On reflection, I think the results are BOGUS. There is no way the i7 will score 65% performance gain (brothers in arms) or 75% (Far Cry 2) over QX9770 unless they are at different clocks.

We know from other benchmarks that HT adds hardly zero value to game FPS, and that i7 core can outcalculate a Core2 at about 10% at the same clocks, same code. (SSE4.x obviously faster, but that is code specific pipeline which I doubt these games are relying on).

There is no way there is another 50% FPS performance from memory bandwidth alone. It just cant be done unless the whole game is relying on memory alone and 100% cache misses. And I dont believe that.

So the results are BOGUS, or the clocks are vastly different, or the X58 is setup is optimum, and the other setups were handicapped in some way.

 

[EarlZ]

It also made me wonder after checking out that review, in other CPU based benchmarks it does not beat a C2Q on the same clock by more than 50%, not even close to 15% in most cases.

 

[Fitseries3]

i want to remind you guys that if you think x58 is too much, wait a few months for the dualcore lynnfield nehalem chips and their respective boards to come out.

the reason x58s are so expensive is due to the 8 layer PCB that it requires to get it all to work.

the p55(i presume) will be on a 6 layer PCB that costs much less then the 8 layer x58s. the dualcore lynnfields will of course be cheaper and more affordable as well.

just something to keep in mind. it seems as though everyone has forgotten about that.

 

[3870x2]

i want to remind you guys that if you think x58 is too much, wait a few months for the dualcore lynnfield nehalem chips and their respective boards to come out.

the reason x58s are so expensive is due to the 8 layer PCB that it requires to get it all to work.

the p55(i presume) will be on a 6 layer PCB that costs much less then the 8 layer x58s. the dualcore lynnfields will of course be cheaper and more affordable as well.

just something to keep in mind. it seems as though everyone has forgotten about that.

not to mention when the technology is the best, that is exactly what you are paying for.

 

[SystemViper]

i want to remind you guys that if you think x58 is too much, wait a few months for the dualcore lynnfield nehalem chips and their respective boards to come out.

the reason x58s are so expensive is due to the 8 layer PCB that it requires to get it all to work.

the p55(i presume) will be on a 6 layer PCB that costs much less then the 8 layer x58s. the dualcore lynnfields will of course be cheaper and more affordable as well.

just something to keep in mind. it seems as though everyone has forgotten about that.

BAM! Also unless you are running benches, or uber big screens, then going to the I7 will not gain you anything since you have not even maxed out your chips you have today.

TO me the I7 is for someone that has cash to play, or wants to show off numbers or is MR. Bling with their 37" Monitors and tri sli or quad crossX. Every new, i mean really new growth in the desktop market takes 6-9 months to just shake out the bugs.

So to me, I would love and dream of getting the I7 but I kill any game i play with my Q9650 and GTX280, and my wokr systems rock too, so I am saving and watching FIT, and will make my decisions next year, when the second gen come out and are more stable, cheaper and have proven track records.....

 

[Bjorn_Of_Iceland]

Is it the CPU or huge memory bandwidth that makes the difference?

Based from my early tweaking days when I got my system, Ive seen no significant FPS gains in SLI from 800Mhz C5 to 1066Mhz C5 (unlocked). I was gaining more FPS when I OC the proc itself (may it be through multiplyer [400*9] or FSB [450*8] the gain was just the same). It may be from the new architecture of how mems are controlled though. But then again I'm not one for conjecture ^^.

AMD is in big trouble, because their "cores" must be rubbish, but the difference has been hidden due to a real win on the memory bandwidth side. But that is now over. They have some serious work to do.

Tis' time for some wicked corporate espionage action for em I guess

 

[SiliconSlick]

Hmm- don't forget the 16000 triple mem

looks like processor scaling really is keeping sli and cfx down. Also, GTX280s choke like a cheap whore on anything above 1920x1200...

Great review link I read a day or two ago but one should remember the triple mem @ 16000 and he has some ocz smokers in there as I recall.

So I say without the mem bandwidth bang the diff would be reduced - you know take the e8400 vs now old quad (since new quad is "oct 4HT")...

Anyway - I think the integrated cpu mem controller and smoking 16000 triple channel is doing the work there - plus the "bus" optimization then that follows - not so much the cpu number crunching power...'cept the e8400 is maxxed.

PS - Would be nice to see xfire 4870 1 gigs huh ... and see what happens - link yet anyone ?
Oh sorry there it is - similar enough with CF 4870 and 4870x2 times two - wow
http://www.guru3d.com/article/core-i7-multigpu-sli-crossfire-game-performance-review/12

 

[SiliconSlick]

I agree somewhat

If it *IS* really only down to memory bandwidth, then that's really interesting.
snip
On reflection, I think the results are BOGUS. There is no way the i7 will score 65% performance gain (brothers in arms) or 75% (Far Cry 2) over QX9770 unless they are at different clocks.
snip
So the results are BOGUS, or the clocks are vastly different, or the X58 is setup is optimum, and the other setups were handicapped in some way.

Yes I go with you somewhat - however they're reaching the maxx on the two older cpu's with the tri and quad setup - so I think it's a combination of (slightly)cpu power and I give the edge to the integrated memory controller- faster triple lanes - and therfore less "bus" talking needed.
So I'd say architecture of the "ramtalk" lanes and triple channel speed...

Without looking again I think you're right about wondering where 75% difference comes from - it's unusual but then so is triple and quad sli and 4xCF with the tops cards

 

[SiliconSlick]

Sorry for 3 posts in a row but you made me wonder

You know, reviewing the link in more detail... especically the OTHER pages earlier in the article, shows some issues in the tests:

1./ 2x SLI GTX260 is FASTER than some charts with 3x SLI GTX280

I suspect there was a problem with 3x SLI setup on the nForce chipset that X58 solves.

So yes, performance is MUCH BETTER on x58 with i7. BUT something isnt working out with the comparison against 3x SLI.

Yes, you really got me curious. This is what he says on the added page to the article:

" Now this is the stuff you only can discover if you have a massive amount of GPU power under the hood. With 2-way SLI GTX 260 the numbers would be much closer to each other as the 2 GPUs do not need a processor of the caliber that Core i7 965 harbors. "

So basically he is saying NOT true in just SLI - but the triple, YES, really needs more.
( he kind of leaves it unanswered because he says the cpu - but we have to include the "ramtalk" speed as part of the cpu now since they integrated it )
I guess it makes sense to me if I think the triple setup needs some SPECIAL "info" (not required in sli) that the integrated mem controller or some oddity in the new core is providing much faster - something that is requested often- so a difference in that bit of info is showing up in fps loss or gains.

 

[lemonadesoda]

Yes, so you basically agree with my observation. You can't square the circle on the data in that review. Something is wrong. Whether the data is wrong, or the hardware is not performing consistently, it should have been obvious to the reviewer that further investigation and diagnotic was necessary. He didnt even notice.

I think we have to wait for a better review/reviewer to get to the bottom of this.

 

[Trizmatic]

[*]BUT, IIRC, cache is still quicker than DDR3 latency and bandwidths, therefore, cache is king, and we should see bigger FPS scaling on different cache sizes on Core 2, but we dont.
[/LIST]

We do.

http://www.tomshardware.com/reviews/cache-size-matter,1709-5.html

 

[Trizmatic]

If you look at their graphs, they mess up the CPU/Board combo a lot. They say they are using the E8400 on the X58 board in a few graphs.
Anyway, thanks for the link. I either missed that part when I read the article or they added it some time after I read it.

 

[lemonadesoda]

We do.

http://www.tomshardware.com/reviews/cache-size-matter,1709-5.html

We dont.

The comment isnt that there is no difference in FPS by cache, but that we would see much bigger scaling. By going from 1MB to 4MB cache there is between 0% and 10% performance gain in gaming in the link you gave.

So how does i7 get 70%. It isnt due to microcode. Since the gains here are 5% as shown in other benchmarks. And it just isnt possible to say this is from the new memory architecture alone. Because if 4MB vs 1MB cache, which is significantly faster than DDR3, can give us an extra 10%, how do we get 70% from DDR3? By extrapolation (which is just an approximation, assuming FPS scales by memory speeds and is unlimited by GPU), if quadrupling the cache (1MB to 4MB) gets you 10% gain, then you would need to quadruple the cache 7x to get 70% performance gain, which would require 4GB cache. ie. the WHOLE GODDAM MEMORY SYSTEM would be operating at cache speeds. Which isnt true. QED.

 

[r9]

looks like processor scaling really is keeping sli and cfx down. Also, GTX280s choke like a cheap whore on anything above 1920x1200...

LOL

 

[r9]

If it *IS* really only down to memory bandwidth, then that's really interesting.

• AMD made the right decision to adopt onboard memory controllers early on.
• AMD is in big trouble, because their "cores" must be rubbish, but the difference has been hidden due to a real win on the memory bandwidth side. But that is now over. They have some serious work to do.
• BUT, IIRC, cache is still quicker than DDR3 latency and bandwidths, therefore, cache is king, and we should see bigger FPS scaling on different cache sizes on Core 2, but we dont.

On reflection, I think the results are BOGUS. There is no way the i7 will score 65% performance gain (brothers in arms) or 75% (Far Cry 2) over QX9770 unless they are at different clocks.

We know from other benchmarks that HT adds hardly zero value to game FPS, and that i7 core can outcalculate a Core2 at about 10% at the same clocks, same code. (SSE4.x obviously faster, but that is code specific pipeline which I doubt these games are relying on).

There is no way there is another 50% FPS performance from memory bandwidth alone. It just cant be done unless the whole game is relying on memory alone and 100% cache misses. And I dont believe that.

So the results are BOGUS, or the clocks are vastly different, or the X58 is setup is optimum, and the other setups were handicapped in some way.

Good point.

 

[r9]

Maybe got to do with the chipset maybe integrated mem controler + triple channel + influence of the x58 + some abducted alien technology known only to intel

 

[Bjorn_Of_Iceland]

Maybe got to do with the chipset maybe integrated mem controler + triple channel + influence of the x58 + some abducted alien technology known only to intel

They have a frozen Megatron in their basement