Intel Xeon 5570: Smashing SAP records.

[Shintai]

I like i7 from architecture perspective. High wattage for the cpu is ok as long as perf/watt benefits and temps do not exceed. I expect temps to be the bigges issue for i7 in server space.

I can guarantee you Xeon 5500 series servers uses LESS power than 5400 series.
And when you can cool 4 and 8 barcelona chips you can cool anything

[justapost]

Yeah, which will appear out of thin air to save them? Or will it rather launch in Q4'09 competing against Beckton (8 cores, 24mb L3, quad channel memory) and Westmere (32nm Nehalem)?

Of course that chip will be available only round 04/09. 8 cores on 45nm? At what frequency?

He is clueless as per Michael S from realwordtech (which BTW I have already linked in my first post) and I would believe a regular RWT poster any day over some random guy bashing anandtech in their comments, which unfortunately is happening all the time.

But the results he posted only differ in cpu utilisation (93% vs. 99%) and the SAPS 5730 vs. 10520. A cause for that lower scoring system could have been some marketing team asking for an crippled k10 result to make k10.5 look better.

More cores will not solve this problem, this is about bandwidth, more cores will likely result in about similar performance then quadcores because the cores are bandwidth starved

Never said istanbul will beat those new xeon's performance wise. Bandwidth wise amd will need maranello to compete here again.

[BakerT]

I can guarantee you Xeon 5500 series servers uses LESS power than 5400 series.
And when you can cool 4 and 8 barcelona chips you can cool anything

I dont get it? Whats funny?

[accord99]

But the results he posted only differ in cpu utilisation (93% vs. 99%) and the SAPS 5730 vs. 10520. A cause for that lower scoring system could have been some marketing team asking for an crippled k10 result to make k10.5 look better.

If you look closer at the two benchmark tests, you'll find the slower one has a response time of ~1s, while the other has a response time of ~2s. This is what Micheal S of Realwordtech forums says about these tests:

Basically, there are two classes of SAP-SD 2-tier submissions - "fast" with response time around 1 second and "throughput-oriented" with response time around 1.6-2 seconds. It seems, average response time above 2 seconds is illegal.
The first class of submissions is far less popular and recently done almost exclusively on x86. All scores in my original post belong to the second class and have average response time very close to 2 sec.

[justapost]

Nopes, have you ever seen those server rack machines. Talking about air cooling, servers have the best and the noisiest (fans fans everywhere, front to back). 2 processors isn't much effort to keep cool, but 4 processors maybe..

I have to manage few 2U servers at work I know their powerfull and noisy fan's . Of course you get the heat easy out off the case. Still have to cool down the server room.

I can guarantee you Xeon 5500 series servers uses LESS power than 5400 series.
And when you can cool 4 and 8 barcelona chips you can cool anything

What do you compare here X5400 or L5400 with 5500? Do you take the gains caused by no longer using fbdimm's into account?
Shanghai's are a differnt story. I did not go barcelona on the server side, neighter 5400. Need a new dual socket server this year so I'm on a lookout. I'd prefere amd because otherwise I'd loose live migration possibility or pefromance i'f id run vm's in an compatible mode.

[Shintai]

I have to manage few 2U servers at work I know their powerfull and noisy fan's . Of course you get the heat easy out off the case. Still have to cool down the server room.

What do you compare here X5400 or L5400 with 5500? Do you take the gains caused by no longer using fbdimm's into account?
Shanghai's are a differnt story. I did not go barcelona on the server side, neighter 5400. Need a new dual socket server this year so I'm on a lookout. I'd prefere amd because otherwise I'd loose live migration possibility or pefromance i'f id run vm's in an compatible mode.

I got 2 5500 series systems. One 3.2Ghz high performance and one 2.26Ghz low voltage (With turbomode). They both use less. And they are compared directly with the previous counter parts from the same OEM. (LV 5400 was 2.33Ghz tho).

Its both due to FB-DIMMs but also chipset. And in performance they are simply a monster. So more or less twice the computational power and alittle lower power consumption. Its regular 1.5V DDR3. Not lowpower 1.35V. They didnt have them yet.

And remember your servers fanspeed can easily cheat due to temperature in the serverroom. They are always overcooling. And noise is different between OEMs. HP servers tend to make alot mroe noise at the same temperature than Dell etc. But ye, if needed the fans can go between 11500 and 14000rpm.

[Donnie27]

My critizism (If I had some because the first was a simple joke) is for how Anandtech presented the facts. i7 is a very fast processor, it is almost like that they have done it to be good for databases. It is a bandwidth monster. But they applaud Intel for having L3 cache and other things used to improve multithreading and in the same article they say that AMD is bad knowing that they was first with L3 and hypertransport (QPI for intel). I don't know how much more you can prove that you isn't neutral.
The facts in the article was way to sparse to do any conclusions what I think. It is no brainer though that i7 will get very good performance running databases. I also think they have some new instructions for parsing data or if some instructions have improved for this (don't remember)

IMHO, it is not about a Fast Processor but a well balanced architecture that had been hobbled by safe but slower features. Hypertransport is born from Alpha EV6 and of all people, RAMBUS. Intel FAB-ed some of the Alpha's. Intel had L3 long before AMD even dreamed about it. Did we forget the P4 Extreme with L3? Give AMD credit for what the really did like X86-64 but please try to stop doing it for stuff they did come up with.

Again, put two Dothan PentiumMs on a QPI and they'd kick mucho-@$$ here. GamePC showed just how much they rocked with a crippled platform. I remember folks partial to the Green Team saying it was fake, just as they said when the first Conroe tests showed up. Many here on this forum Doubted the first i7 reviews when others thought they were s/low. This parallel monster gets meaner as its L3 grows, memory controller/s get larger/faster and yes, QPI speeds up. The biggest problem AMD has is that Intel

[gosh]

Donnie27 I don't agree with your history lesson.

On servers i7 need to get the power down

[Movieman]

IMHO, it is not about a Fast Processor but a well balanced architecture that had been hobbled by safe but slower features. Hypertransport is born from Alpha EV6 and of all people, RAMBUS. Intel FAB-ed some of the Alpha's. Intel had L3 long before AMD even dreamed about it. Did we forget the P4 Extreme with L3? Give AMD credit for what the really did like X86-64 but please try to stop doing it for stuff they did come up with.

Again, put two Dothan PentiumMs on a QPI and they'd kick mucho-@$$ here. GamePC showed just how much they rocked with a crippled platform. I remember folks partial to the Green Team saying it was fake, just as they said when the first Conroe tests showed up. Many here on this forum Doubted the first i7 reviews when others thought they were s/low. This parallel monster gets meaner as its L3 grows, memory controller/s get larger/faster and yes, QPI speeds up. The biggest problem AMD has is that Intel

I have a dual socket system with essentially 2 yonah(dualcore Dothan) laptop chips in it.
4 cores total at 1995/2mb/667.
Computational power equal to a Q6600 at the same mhz and draws 117w at 100% load,85w at idle.
Actually using it to post now as my elec is out and on a generator and this is the lowest drawing system in the house!

[justapost]

If you look closer at the two benchmark tests, you'll find the slower one has a response time of ~1s, while the other has a response time of ~2s. This is what Micheal S of Realwordtech forums says about these tests:

Thanks for the details. So those benchmarks with the low thruput where intentionally optimized for fast response times. Not so unrealistic if workload are people whom have to feed the SAP system with data thru dialog input.

I got 2 5500 series systems. One 3.2Ghz high performance and one 2.26Ghz low voltage (With turbomode). They both use less. And they are compared directly with the previous counter parts from the same OEM. (LV 5400 was 2.33Ghz tho).

Its both due to FB-DIMMs but also chipset. And in performance they are simply a monster. So more or less twice the computational power and alittle lower power consumption. Its regular 1.5V DDR3. Not lowpower 1.35V. They didnt have them yet.

2.26GHz low voltage version would be an option here. You mean lower in idle?
Performance is not that much an issue, response time is more important.

And remember your servers fanspeed can easily cheat due to temperature in the serverroom. They are always overcooling. And noise is different between OEMs. HP servers tend to make alot mroe noise at the same temperature than Dell etc. But ye, if needed the fans can go between 11500 and 14000rpm.

I'm more concerned about cooling. I expect it is more efficient to avoid heat in above scenario.

[Shintai]

Donnie27 I don't agree with your history lesson.

On servers i7 need to get the power down

You are confusing. Its already lower. And specially on a performance scale its massively lower. And no, Shanghai doesnt beat it.

[Shintai]

2.26GHz low voltage version would be an option here. You mean lower in idle? Performance is not that much an issue, response time is more important.

I'm more concerned about cooling. I expect it is more efficient to avoid heat in above scenario.

Mean lower in idle? No...lower in everything. 100%load and idle.

Cooling is not an issue. If it was P4 and Barcelona servers should explode.

5500 series servers generate LESS heat than 5400 series.

[Ghostbuster]

@gosh, the Hypertransport and DDR IMC was derived from Alpha's bus designs which Intel bought over later. In fact AMD got the EV6/EV7 license FREE from (guess who?) Intel.

The Gallatin core had an integrated L3 cache (plus Northwoord core), and was considered the fastest processor during its (very short) reign. It can even go head to head against Athlon 3200+ at that time and still win. However it was also the most expensive around (having the "Extreme Edition" tag). This is an extremely rare chip.

The Dothan was an entirely different design from the Netburst. Its more towards higher IPC and low power.

[gosh]

Please read carefully!
I haven't said that AMD has invented L3 cache, hypertransport etc.
What I said was that the article focused on how good i7 while they informed how bad AMD is. AMD had L3 cache on die before i7, AMD had hypertransport before i7 QPI.
Do you understand or is this to complicated?

[Clairvoyant129]

Please read carefully!
I haven't said that AMD has invented L3 cache, hypertransport etc.
What I said was that the article focused on how good i7 while they informed how bad AMD is. AMD had L3 cache on die before i7, AMD had hypertransport before i7 QPI.
Do you understand or is this to complicated?

But it did perform badly, dual socket systems spanking quad socket Opterons.

[gosh]

But it did perform badly, dual socket systems spanking quad socket Opterons.

And you need more information before you can draw any conclusions.

Why did Intel let go of the NDA? Doesn't that seem a bit desperate?

[Ghostbuster]

And you need more information before you can draw any conclusions.

Why did Intel let go of the NDA? Doesn't that seem a bit desperate?

So how do you draw your conclusions that Intel "seem a bit desparate"? Don't forget the SPECint scores also...

[savantu]

News at 11 : 2S Nehalem beats 4S Shanghai in SAP-SD

2s/8c Nehalem (2.93 GHz) - 4995 users
2s/8c Shanghai (2.7 GHz) - 2730 users
4s/16c Shanghai (2.7 GHz) - 4386 users

Basically , one Nehalem thread is faster than a Shanghai core in this particular benchmark.

http://www.sap.com/solutions/benchmark/sd2tier.epx

Dunnington fairs pretty poorly at 4400 users per 24 cores.1066MHz FSB and FBDIMM 667 are real show stoppers.

[gosh]

So how do you draw your conclusions that Intel "seem a bit desparate"? Don't forget the SPECint scores also...

When will you be able to buy these servers with two sockets i7?

[Oliverda]

When will you be able to buy these servers with two sockets i7?

After Intel resolve some of the C0 stepping erratas and the cache coherency issue.

[Shintai]

When will you be able to buy these servers with two sockets i7?

Next month according to OEMs.

[gosh]

Next month according to OEMs.

If it is next month then it is ok, not desperate

[Donnie27]

I have a dual socket system with essentially 2 yonah(dualcore Dothan) laptop chips in it.
4 cores total at 1995/2mb/667.
Computational power equal to a Q6600 at the same mhz and draws 117w at 100% load,85w at idle.
Actually using it to post now as my elec is out and on a generator and this is the lowest drawing system in the house!

Yes but I meant put two of them with QPI + IMC and they'd easily kick mucho-asseo! I'm not comparing them on cripple platforms they had back in the day.

[Macadamia]

More cores will not solve this problem, this is about bandwidth, more cores will likely result in about similar performance then quadcores because the cores are bandwidth starved

This is why Dunnington shows a near-1.5x scaling compared to Tigerton.

[Movieman]

Yes but I meant put two of them with QPI + IMC and they'd easily kick mucho-asseo! I'm not comparing them on cripple platforms they had back in the day.

Yup, I understood. Just saying they are a sweet little system even though on a crippled platform.