Intel Xeon 5570: Smashing SAP records SCOOP

[Idontcare]

The SAP numbers are absolutely astonishing, as Intel's dual socket is able to outperform quad socket opteron machines.

Based on the scaling of Barcelona, we speculate that a quad Shanghai at 2.7 GHz would obtain the performance of the Dual Xeon 5570 w/o HT.The new Xeon 5570 outperforms the "old" 5450 by 119%!!!

http://it.anandtech.com/weblog/showpost.aspx?i=532

 

[JackyP]

How important is SAP-SD in the server-world? What would this mean to AMD's server share if Nehalem is twice as fast per clock in this application?
Are there any ways to "cheat" on SAP-SD? Does the benchmark fit into RAM or does it access the hard drives too? (maybe they used SSDs as proposed by someone on the aces-board?)

 

[Nemesis 1]

Its just amazing. Amd gets 45nm to run at 6ghz. Intel fanbois start yelling cheat.

Better this post not been made. The cheat thing will just keep poping up. Latter this year when we see mature M/Bs and new steppings . Its only going to get better and better on the desktop. THan if Hydra delivers as promised. THe IC7 will show its true value.

 

[ajaidevsingh]

Read the whole article the numbers are by HP, etc not Anand tech:-

"These numbers are so high, that we checked and checked again. The database used is the same (SQL Server 2005), so unless there is some incredible tuning parameter that HP and FS have discovered and that we have yet to hear about, that is not it.

At this point we have no idea how it is possible that a 3 GHz Nehalem outperforms the latest Opteron by a margin as high as 80% and more."

 

[ajaidevsingh]

Originally posted by: JackyP
How important is SAP-SD in the server-world? What would this mean to AMD's server share if Nehalem is twice as fast per clock in this application?
Are there any ways to "cheat" on SAP-SD? Does the benchmark fit into RAM or does it access the hard drives too? (maybe they used SSDs as proposed by someone on the aces-board?)

It is basically an bench app for client to server and back visualizer. Sync latency can be artificially lowered and that is one of the corner stones of this app.

Have tried the out when it was released and i got a score of around 3000 on an rather new Xeon machine. So the new xeons work 6-7 times faster than the server at my collage wow that is unreal.

 

[aigomorla]

uhhhh wow..

AMD is looking good..

The top intel is a 24 core monster incase any of you guys want to know...

the X7460 is a hex core chip. x 4 LOL

IDC if i follow those charts carefully, the chips i have should rank with that 24 core monster with HT enabled..

LOL..

 

[magreen]

I was waiting for this, given that all of Nehalem's features seem to be designed for the server space, and seeing that we didn't see massive improvements on the desktop. I pretty much trusted there'd be some big guns to this thing when the cover gets lifted.

So I'm not too surprised to see big server performance numbers.

 

[ajaidevsingh]

Seems i was correct about the HT "fake core"

Update (a few hours later): It seems that the SAP page was wrong about HT. It reported 8 threads on 8 cores on the Fujitsu Siemens Primergy Server. The certification page says otherwise: 16 threads on 8 cores. So hyperthreading (SMT) plays probably an important role in this benchmark as the SAP application has very low IPC and is very parallel. So this completely annihilating performance comes from combining a wide superscalar CPU with an excellent Simultaneous Multithreading implementation. Hats off to the Intel engineers...

 

[Phynaz]

How can an application have low IPC?

 

[magreen]

Originally posted by: ajaidevsingh
Seems i was correct about the HT "fake core"

Update (a few hours later): It seems that the SAP page was wrong about HT. It reported 8 threads on 8 cores on the Fujitsu Siemens Primergy Server. The certification page says otherwise: 16 threads on 8 cores. So hyperthreading (SMT) plays probably an important role in this benchmark as the SAP application has very low IPC and is very parallel. So this completely annihilating performance comes from combining a wide superscalar CPU with an excellent Simultaneous Multithreading implementation. Hats off to the Intel engineers...

Can someone translate this for me? What's changed given this update? Just that they didn't think it was utilizing HT and now they see it is? So then why is the HT disabled number in the chart almost as high as the HT enabled result?

 

[Idontcare]

Originally posted by: magreen
So then why is the HT disabled number in the chart almost as high as the HT enabled result?

The data are correct, the chart label for the data is incorrect. In both cases HT was enabled, hence the observation of the data being similar to each other.

 

[magreen]

Originally posted by: Idontcare

Originally posted by: magreen
So then why is the HT disabled number in the chart almost as high as the HT enabled result?

The data are correct, the chart label for the data is incorrect. In both cases HT was enabled, hence the observation of the data being similar to each other.

Ok I get it now, thanks. And so what new factor makes Johan claim that SMT is responsible for the result now? The only thing that makes sense is if he would have originally thought SMT was responsible for the high numbers, but then ruled out SMT by the high numbers the Fujitsu-Siemens system posted w/o HT... and now that he sees HT was enabled, he goes back to his original theory that SMT would be responsible for such a high result? Am i on the right track here?

 

[Idontcare]

Originally posted by: magreen

Originally posted by: Idontcare

Originally posted by: magreen
So then why is the HT disabled number in the chart almost as high as the HT enabled result?

The data are correct, the chart label for the data is incorrect. In both cases HT was enabled, hence the observation of the data being similar to each other.

Ok I get it now, thanks. And so what new factor makes Johan claim that SMT is responsible for the result now? The only thing that makes sense is if he would have originally thought SMT was responsible for the high numbers, but then ruled out SMT by the high numbers the Fujitsu-Siemens system posted w/o HT... and now that he sees HT was enabled, he goes back to his original theory that SMT would be responsible for such a high result? Am i on the right track here?

I haven't checked to see if he edited the blog or not. Hey, it was free data mining provided by Johan, we got what we paid for. No harm no foul.

 

[dmens]

Originally posted by: Phynaz
How can an application have low IPC?

i think he meant there's a lot of waiting from memory. SMT is good in that situation.

 

[Dadofamunky]

I sure wish the HP engineers could weigh in with their observations. I'm REALLY curious about this. If they hold up, the numbers are just devastating. Have the AT guys thought about getting in touch with them? I can't imagine it'd be that hard. In my fairly deep experience, Silicon Valley engineers are pretty gregarious and happy to share information like this.

 

[BLaber]

The real competition is when Amd's Shanghai gets ddr3.0 and ht3 treatment in Q4 09 , along with that an Istanbul-6 core to compete with Nehalem on server 2p , 4p platforms, just can't wait..

 

[Idontcare]

Originally posted by: BLaber
The real competition is...

In Q4'09 we are talking about Nehalem-EX, aka Beckton, with its 8 cores and 16 threads plus 24MB L3$ and 4x QPI for 4P boards.

Originally posted by: BLaber
when Amd's Shanghai gets ddr3.0 and ht3 treatment in Q4 09

What's that going to add, maybe a 5% IPC bump?

Originally posted by: BLaber
along with that an Istanbul-6 core to compete with Nehalem on server 2p , 4p platforms, just can't wait..

That's good, but IF it comes out on that timetable it will be competing with 8-core Beckton.

And then 3-6 months later Intel's 32nm server chips hit resellers.

IMO AMD's next opportunity to stand toe-to-toe with Intel won't come until they release Sao Paolo and Magny-Cours.

Until then the story will be gross margin erosion across the Opteron line if AMD wants volume sales that exceed the usual upgrade cycle fanfare.

Unless Intel takes the high-road and pegs Nehalem Xeon prices in the $3k territory.

(which they easily could, but they may have trouble resisting the desire to pile more hurt on AMD, a self-defeating move in the end IMO as DOJ moves in and dismantles Intel post AMD demise)

The conclusion is clear. AMD needs to lobby congress for a bailout package.

 

[ajaidevsingh]

For people who did not understand this app and are confused as to why are the scores that high for the i7.

i7 is basically a radioactive monster merger of P4 with HT and a Athlon 64. HT shows that the cpu has extra cores now mind you older xeons did not have this. Even at the time of the P4 some benchs were artificially pumped due to HT and that is the samecase here.

See the results of GTA4 bench you will not the low i7 CPU utilization that is not because the CPU is too good for the game its because of the HT. Now also this bench is very core sensitive as in more cores more score "parallel execution" and HT allows that so that means tough a dual i7 server setup has 8 cores it can play with 16 Threads as to a quad set up of Denebs which has total of 16 cores and can play around with 16 threads.

 

[IntelUser2000]

How can an application have low IPC?

Basically, there are some applications that can't be parallelized much by Instruction-Level parallelism, cause the fundamental code just doesn't allow it. Remember the new CPUs have the capability to decode, execute and retire 4 instructions per cycle, like the Core 2 Duo. Pentium III/Athlon/Pentium M all have 3.

In server applications like SAP and TPC-C, the average IPC for the application barely exceeds 1. The CPU runs most of the code at 1-1.5 IPC.

In these applications, Simultaneous Multi-Threading can help immensely. SMT(HT is Intel's name) is a way of utilizing those execution units which are under utilized(because of 1 IPC out of possible 3 or 4).

30-40% improvements are possible for SAP.

 

[Idontcare]

Originally posted by: ajaidevsingh
i7 is basically a radioactive monster merger of P4 with HT and a Athlon 64.

A professional assessment based on your microarchitecture engineering background?

Originally posted by: ajaidevsingh
Even at the time of the P4 some benchs were artificially pumped due to HT and that is the samecase here.

How can performance be artificial? It seems pretty obvious to me that if you use SAP then HT on Nehalem is going to be a benefit. Not an artificial benefit.

Originally posted by: ajaidevsingh
See the results of GTA4 bench you will not the low i7 CPU utilization that is not because the CPU is too good for the game its because of the HT. Now also this bench is very core sensitive as in more cores more score "parallel execution" and HT allows that so that means tough a dual i7 server setup has 8 cores it can play with 16 Threads as to a quad set up of Denebs which has total of 16 cores and can play around with 16 threads.

If the fps is above the minimum fps threshold for human detection then who cares if CPU utilization is 5% or 95%? What should matter at that point is power consumption while playing the game...which I have yet to see the i7 deliver inferior performance/watt in gaming benches thus far.

Hopefully when the PhII reviews are rolling out someone does it some justice and captures average power consumption to go along with those average FPS numbers per game so we can see what advantages PhII delivers in this regard.

 

[magreen]

Originally posted by: Idontcare

Originally posted by: magreen

Originally posted by: Idontcare

Originally posted by: magreen
So then why is the HT disabled number in the chart almost as high as the HT enabled result?

The data are correct, the chart label for the data is incorrect. In both cases HT was enabled, hence the observation of the data being similar to each other.

Ok I get it now, thanks. And so what new factor makes Johan claim that SMT is responsible for the result now? The only thing that makes sense is if he would have originally thought SMT was responsible for the high numbers, but then ruled out SMT by the high numbers the Fujitsu-Siemens system posted w/o HT... and now that he sees HT was enabled, he goes back to his original theory that SMT would be responsible for such a high result? Am i on the right track here?

I haven't checked to see if he edited the blog or not. Hey, it was free data mining provided by Johan, we got what we paid for. No harm no foul.

Hey, I'm not criticizing Johan. I'm assuming that he knows more than me, and then I'm asking what he's basing his conclusions on.

 

[IntelUser2000]

Those who yet doubt Intel's claims SMT will be much better on Nehalem than P4 based CPUs will get to know soon.

Keep saying, P4 lost a lot of HT's opportunity since it had a architectural flaw. Nehalem doesn't.

The surprise Intel brought for Core 2 on desktops is gonna be brought with Nehalem and SpecRATE and SAP will be the first few.

Get ready to see what Nehalem can do.

 

[ajaidevsingh]

Originally posted by: Idontcare

Originally posted by: ajaidevsingh
i7 is basically a radioactive monster merger of P4 with HT and a Athlon 64.

A professional assessment based on your microarchitecture engineering background?

Originally posted by: ajaidevsingh
Even at the time of the P4 some benchs were artificially pumped due to HT and that is the samecase here.

How can performance be artificial? It seems pretty obvious to me that if you use SAP then HT on Nehalem is going to be a benefit. Not an artificial benefit.

Originally posted by: ajaidevsingh
See the results of GTA4 bench you will not the low i7 CPU utilization that is not because the CPU is too good for the game its because of the HT. Now also this bench is very core sensitive as in more cores more score "parallel execution" and HT allows that so that means tough a dual i7 server setup has 8 cores it can play with 16 Threads as to a quad set up of Denebs which has total of 16 cores and can play around with 16 threads.

If the fps is above the minimum fps threshold for human detection then who cares if CPU utilization is 5% or 95%? What should matter at that point is power consumption while playing the game...which I have yet to see the i7 deliver inferior performance/watt in gaming benches thus far.

Hopefully when the PhII reviews are rolling out someone does it some justice and captures average power consumption to go along with those average FPS numbers per game so we can see what advantages PhII delivers in this regard.

I was trying to explain the integrated components "AMD did that with Athlon 64" and the HT "Intel introed it with P4"

Performance can be artificial in benchmarks pls read the UPDATE made in the Anadtech link.

GTA4 was an example for the architecture and how it can fool the software into thinking there are 8 cores instead of 4 with the use of HT. As it stands GTA4 is not the most intelligent game thinking my Q6600 is a PIII xeon :roll:

 

[ajaidevsingh]

Originally posted by: IntelUser2000
Those who yet doubt Intel's claims SMT will be much better on Nehalem than P4 based CPUs will get to know soon.

Keep saying, P4 lost a lot of HT's opportunity since it had a architectural flaw. Nehalem doesn't.

The surprise Intel brought for Core 2 on desktops is gonna be brought with Nehalem and SpecRATE and SAP will be the first few.

Get ready to see what Nehalem can do.

I agree the Nehalem architecture was more focused on HPC and server market than Desktops but do you think that a Nehalem chip can double the performance of the last best by Intel..!!

SMT has gone the rite way in the server and HPC market and will show its potential when apps take full advantage of this feature. I wanted to see Virtulization score and HPC apps those would do wonders with more raw speed!!:camera:

 

[aigomorla]

Originally posted by: IntelUser2000

Get ready to see what Nehalem can do.

playing with neha-ep already.... lol...

I need a dual cpu tylersberg board now...