First Shanghai benchmark results.

[qurious63ss]

Spec_power benchmarks turns up and it looks like shanghai is much more efficient then harpertown.

Disclaimer: AMD submitted the results for the Intel machine so....it may be have an effect on the results. Let the number spinning commence.


http://www.theinquirer.net/gb/inquir...ghai-benchmark

[demonkevy666]

where's the comparison of Barcelona ?

[Hornet331]

why the amd rigs only run 2 instances and intels 4?

why are the amd rings using a 500W PSU and the intel a 700W, especial when max consumpton is only 13W appart?

even the inq noticed this...

[informal]

why the amd rigs only run 2 instances and intels 4?

why are the amd rings using a 500W PSU and the intel a 700W, especial when max consumpton is only 13W appart?

even the inq noticed this...

It's seems as a common practice with all the systems based on Xeons(maybe because of the MCM nature of the design).

Each JVM instance was affinitized to a pair of cores.

Anyhow,the compared systems should use same amount of memory(16GB for example).
Performance wise Shanghai is very strong in this test(pure bench numbers: 2P Sh@ 2.7Ghz scores 338K),while power draw could be a bit better i suppose.Still these are very early submissions.

[JumpingJack]

why the amd rigs only run 2 instances and intels 4?

why are the amd rings using a 500W PSU and the intel a 700W, especial when max consumpton is only 13W appart?

even the inq noticed this...

This is where analyzing power efficiency numbers can be tough. AMD is going to do what any company will do to promote a product. They configured their system, then configured the Intel system in a similar matter.

Intel's Achilles heal is FBDIMMs, any power efficiency in the CPU is lost in the memory. If you scan through all the other benchmarks, there are intel systems scoring 1000+ but do not utilized 16 Gigs of 8 sticks of 2 Gigs apiece. So take the AMD submission with a grain of salt.

But Informal made the right observation -- the full load ssj_ops for Shanghai vs equivalently clocked Harpertown is higher, even in cases of 2 instant JVMs ... unfortunately, since there is no Barcelona SpecPower benchmark submitted, it is not possible to validate 20% claims ... but the raw performance showing against Harpertown is impressive. i.e. http://www.spec.org/power_ssj2008/re...519-00056.html
It would appear a 2.7 GHz Shanghai in two sockets is about the same as a 3.0 GHz Harpertown.

[informal]

Yep,good post Jack.That Harper system is using a lot power though,20% more,while performing on par with 2.7Ghz Sh. Memory config. is comparable(4x4GB).

What AMD could do is use single socket systems(will there be any available soon?) and post very high scores(for single socket) with low power draw.This way they can boost the "overall" score a lot,but with a lot less ssj_ops.

[Macadamia]

I have to say though, that 2P downwards isn't their forte.
Sure, Shanghai IS going to get bought a LOT simply because it is a great slot in upgrade without the need for any platform changes or validation, but sooner or later Gainestown will come and probably make quite a nice lead even clock for clock (power wise I have doubts though, since Nehalem's process was more oriented towards an initial high "flatrate" and later lower voltage required for raising clocks)

4P will be a serious fightback, since Beckton is slow as usual.

[JumpingJack]

Yep,good post Jack.That Harper system is using a lot power though,20% more,while performing on par with 2.7Ghz Sh. Memory config. is comparable(4x4GB).

What AMD could do is use single socket systems(will there be any available soon?) and post very high scores(for single socket) with low power draw.This way they can boost the "overall" score a lot,but with a lot less ssj_ops.

Yeah ... I am always beating my head against a wall when people argue about power efficiency ... it is very problematic, there is not set standard how to do it.

Sure SpecPower is a start, but even this has massive problems, and is too 'tunnel - visioned', loading it down in just one scenario.

One major problem is that power efficiency is a 2 vector metric, a small change in the denominator can lead to huge swings in the result ... I think a better way of enunciating the metric would be to normalize the power (the denominator) with costs or total energy consumed over a typical usage.

So a processor consumes 10% less power, 10% less power translates into what costs savings in the long run vs a processor that runs faster but uses 10% more though get's to idle quicker...

It is just to convoluted to be of any good really.

[duploxxx]

why the amd rigs only run 2 instances and intels 4?

why are the amd rings using a 500W PSU and the intel a 700W, especial when max consumpton is only 13W appart?

even the inq noticed this...

you can never compare all these results, therefore the overall clasification is too wide for own creations... and there are lots of other benchmarks done with other psu and 2-4 threads

for example:

Intel Corp. Super Micro Computer, Inc. Supermicro 6025B-TR+ (Intel Xeon processor E5345) BEA JRockit(R) 1.6.0_02 (build P27.4.0-10-90053-1.6.0_02-20071009-1827-windows-x86_64, compiled mode) 2 419 Text HTML
Intel Corp. Super Micro Computer, Inc. Supermicro 6025B-TR+ (Intel Xeon processor E5345) BEA JRockit(R) 1.6.0_02 (build P27.4.0-10-90053-1.6.0_02-20071009-1827-windows-x86_64, compiled mode) 4 435 Text HTML

This is where analyzing power efficiency numbers can be tough. AMD is going to do what any company will do to promote a product. They configured their system, then configured the Intel system in a similar matter.

Intel's Achilles heal is FBDIMMs, any power efficiency in the CPU is lost in the memory. If you scan through all the other benchmarks, there are intel systems scoring 1000+ but do not utilized 16 Gigs of 8 sticks of 2 Gigs apiece. So take the AMD submission with a grain of salt.

But Informal made the right observation -- the full load ssj_ops for Shanghai vs equivalently clocked Harpertown is higher, even in cases of 2 instant JVMs ... unfortunately, since there is no Barcelona SpecPower benchmark submitted, it is not possible to validate 20% claims ... but the raw performance showing against Harpertown is impressive. i.e. http://www.spec.org/power_ssj2008/re...519-00056.html
It would appear a 2.7 GHz Shanghai in two sockets is about the same as a 3.0 GHz Harpertown.

the difference in idle and load is the same, more or less 60Watts, both work with 4 dimms so its not only the fbdimm consuming the difference, still looks like powerdraw on shanghai improved a lot as long as they stay under certain frequency.

[qurious63ss]

It looks like at 2.7ghz the shanghai uses 264W with a perf to power ratio of 860, while the harpertown running at 2.66ghz uses only 161W with a perf to power ratio of 1284. What's going on with shanghais power numbers?

[duploxxx]

It looks like at 2.7ghz the shanghai uses 264W with a perf to power ratio of 860, while the harpertown running at 2.66ghz uses only 161W with a perf to power ratio of 1284. What's going on with shanghais power numbers?

that's the difference between a Low voltage cpu with lower bus speed etc and a normal cpu... if you check the ssj_ops they are also way lower then the E5472 and the 2384 but the power is also difference, btw it has only 2 fbdimms, so you can already add +10 more watts when stressed for the ram.

[gOJDO]

When I red the title I though I'm going to see some numbers from an individual. But it's coming from AMD it self...I'd rather believe in the opposite of what they claim.

[Shintai]

Last time AMD also claimed K10 was 40 and 50% faster than Core 2. Plus this coupled with theinq. And being spec numbers. No thanks..

And funny to see the Intel system by AMD scoring so low compared to what others archieve.

Sounds like a rerun of the Barcelona...

[informal]

Shintai,AMD "Configured" system (the intel based one) is not under performing in pure performance,it is on par with other results.What is different is the power draw since the systems are equipped with the same amount of RAM memory(16GB).
In pure performance,like Jack said,Shanghai @2.7Ghz is on the level of Harpertown at 3GHz,in 2P.In 4P it should be even greater difference,but we have no data for that yet.

[Helmore]

Shintai,AMD "Configured" system (the intel based one) is not under performing in pure performance,it is on par with other results.What is different is the power draw since the systems are equipped with the same amount of RAM memory(16GB).
In pure performance,like Jack said,Shanghai @2.7Ghz is on the level of Harpertown at 3GHz,in 2P.In 4P it should be even greater difference,but we have no data for that yet.

Performance in what scenario? What would it be like for video encoding and 3d rendering? Or maybe even things like games? It may be like that for certain performance figures, but I don't think this will really apply to most users here. I mean, what do those Spec numbers mean to the average user here?

[informal]

Performance in what scenario? What would it be like for video encoding and 3d rendering? Or maybe even things like games? It may be like that for certain performance figures, but I don't think this will really apply to most users here. I mean, what do those Spec numbers mean to the average user here?

If you haven't noticed,we are all speaking about server workloads here,not desktop.Shanghai is a server part,after all... As for desktop,we simply don't know what kind of perf. over even Agena new 45nm chip will offer.We simply have to wait a bit longer to see.What it will bring to desktop for sure is 3 things: higher stock clocks-2.8/3Ghz in Dec,much lower power draw and much higher OCing.These 3 factors are equally important as IPC is,so Deneb if priced right will be a very good desktop chip.Oh and you can guess even now it's not going to be slower than Agena...

[Rammsteiner]

Quite interesting.

However, I'd rather take those results not too serious. Both from Shanghai as Harpertown. Marketing is most of the times playing a big role in such 'previews' made my the producer its self.

But well, not too long before we can see some real benchies

[Shintai]

If you haven't noticed,we are all speaking about server workloads here,not desktop.Shanghai is a server part,after all... As for desktop,we simply don't know what kind of perf. over even Agena new 45nm chip will offer.We simply have to wait a bit longer to see.What it will bring to desktop for sure is 3 things: higher stock clocks-2.8/3Ghz in Dec,much lower power draw and much higher OCing.These 3 factors are equally important as IPC is,so Deneb if priced right will be a very good desktop chip.Oh and you can guess even now it's not going to be slower than Agena...

The initial benchmark addresses only one subset of server workloads: the performance of server side Java. It exercises the CPUs, caches, memory hierarchy and the scalability of shared memory processors (SMPs) as well as the implementations of the JVM (Java Virtual Machine), JIT (Just-In-Time) compiler, garbage collection, threads and some aspects of the operating system.

Not much serverload there.

http://www.spec.org/power_ssj2008/re...r_ssj2008.html

And funny how the AMD Intel system doesnt match the others.

No wonder they been so quiet about the 45nm K10s. Nothing to see.

[Rammsteiner]

Not much serverload there.

http://www.spec.org/power_ssj2008/re...r_ssj2008.html

And funny how the AMD Intel system doesnt match the others.

No wonder they been so quiet about the 45nm K10s. Nothing to see.

Now I wouldnt say that either.

I cant remember, but they said with Barcelona I think it was 40% faster than, dont know what. In a lot of benches it actually wasnt, but in one certain it actually was. Now it's not really good to say it's upto 40% faster, but it ain't a lie neither.

I do hope this time AMD did it better than that, and I do think Shanghai/Deneb is a very good improvement from their current products, but as well compared to Intel's offers. How good, we'll have to see untill unbiased reviews are around.

[hollo]

Intel chips certainly seem to be the big winners as far as this particular benchmark is concerned, with Power Leader running Chipzilla’s L5430 Xeons and scoring 1135. IBM also pulls two very high scores from its hat with Intel’s Xeon X3360 2.83GHz (1064, 1054), while NEC also posts a high score of 1010 on a Xeon L5420 2.5GHz.

But there’s also something a bit suspect going on in these benchmarks, something which makes us think AMD might be up to some shenanigans.

AMD submitted its own results for Intel Xeon L5420s at 2.5GHz and somehow managed to come up with a score of only 561.

this = shenanigans

[Rammsteiner]

lame

You act like this is the first time you see this happening?

Nothing new here really

[AliG]

interesting, I wonder how legit this numbers are compared to the average server system if what amd says is true (granted we have to take some of it with a grain of salt)

[Shintai]

· Each server uses the same JVM version
· Each server uses the same JVM command-line options

Yet they use different amount of instances.

And still so different from anyone else....sure!

And since when is a normal server configuration 16GB? Plus you would most likely use 4x4GB and not 8x2GB.

I put this along the 40% faster and 3Ghz Phenoms.

They didnt learn from it last time it seems.
http://uk.youtube.com/watch?v=G_n3wv...eature=related
http://uk.youtube.com/watch?v=R7EZmY...eature=related
http://uk.youtube.com/watch?v=6ipfxI...eature=related

[JumpingJack]

this = shenanigans

Well, regardless of how these scores are made or submitted, anytime a company does a coca cola vs pepsi test by configuring the competitors system themselves one should induce a degree of skepticism ... AMD often does this, I can only recall Intel doing their own AMD configuration compare and I nearly unhinged my jaw in disbelief -- i.e. not the result but rather -- rarely does Intel even acknowledge publicly AMD performance or even existence, but Conroe/Woodcrest filled the IDF and the Intel benchmark website with AMD comparisons.

So when Intel did the Conroe coca cola v pepsi demo, the entire conglomerate (HW journalists, forum enthusiasts, and the like) were all 'take it with a grain of salt.' ... I think it more prudent to weigh the numbers accordingly, note the interesting performance (not performance/watt, just raw performance), then wait for SPEC database to fill with more independent submissions with more configurations.

Bottom line -- Shanghai is shaping up much better than Barcelona did ... forget the AMD sponsored comparison all together.

[JumpingJack]

Performance in what scenario? What would it be like for video encoding and 3d rendering? Or maybe even things like games? It may be like that for certain performance figures, but I don't think this will really apply to most users here. I mean, what do those Spec numbers mean to the average user here?

These are purely server related tasks, the applicability of these type of measurements are completely irrelevant for client side applications.

For servers, AMD has done a good job with Shanghai looking at the raw performance and not performance/watt ... AMD's superior interconnect platform over the archaic front side bus is huge advantage, they don't choke up as they add more and more cores and sockets ... hence, they claim to fame with scaling. In single socket desktop PCs the rules are completely different, Deneb will shed a lot of light on what they managed to complete with architectural tweaks during the shrink ... that should follow Shanghai shortly.