Feeds

back to article IBM chops high-end Power6 server tags

IBM has been mum about when its high-end servers based on the eight-core Power7 processor are going to come to market. With bigger x64 and Itanium boxes around the corner and the 256-core Power 795 machines using Power7s not expected until later this year, though, Big Blue has to do something to get at least some customers to …

COMMENTS

This topic is closed for new posts.
Thumb Up

its not about list price

Those systems are discounted and saved us over $100K per processor for our oracle license stack......the funny thing is IBM could have charged us 4x the cost and we would have still bought them to avoid the oracle licenses we would have had to buy....we have become a huge Power fan because of Oracle's insane pricing. Between vmware and powervm we have drastically reduced the datacenter.

cheers from Dallas

0
0
Joke

Where is Matty ?

..when you need him to decry the dangerously high-clocked contraption from IBM ?

Couldn't the high clock rate generate a black hole and swallow the earth ? Safety is on the side of Itanic !

0
0
Silver badge
Happy

Sunshiner cross-over?

Dear oh dear, poor ickle jlocke has given up championing SPARC and started singing the Power song. Ah well, just goes to show even he's not quite as thick as he comes across, even he realises SPARC in any form is a lost cause. Not surprised he's jumped for the safety of "big boy" IBM and Power after what must have been for him the very shocking Sunset. Anyway, on with the fun!

"....Couldn't the high clock rate generate a black hole and swallow the earth ?...." I suppose that, after the years of Sunshine and SPARC vapourware, it's not surprising that your grasp on reality is a little weak. I hope you don't post under your real name as your employer (even if it is the local McDonalds) probably wouldn't be too impressed with your deluded ramblings.

"....Safety is on the side of Itanic !" Well, safety is on the side of Itanium, Power, Xeon and Opteron users, as none of those chips are in doubt, whereas just about anything left from the Sun carcass is vulnerable to Larry's demand for profits. Maybe I should take time to explain the word "profit" seeing as most of today's Sunshiners have never known a time when Sun actually made a profit? Then again, seeing as most of them have serious problems comprehending even the most simplistic business, economic or technical facts, it would probably be time wasted!

IBM's price-cuts are very understandable - they need to shift existing stock as most customers will understandably want to wait for Power7, and those holdouts could put a dent in IBM's revenue figures for the next two quarters. I'm also pretty sure it's becasue the DDR2-equipped Power6 doesn't stand much of a chance against DDR3-equipped Tukzilla, and seeing as hp might get their Tukzilla top-end kit out before the Power7 top-end arrives IBM is going to need every help it can get to stop those more lucrative top-end deals going to hp.

0
1
Go

Tukzilla dates

Hi Matt,

HP's April 27 announce will only be one/two/four wide blade offerings.

The SX3000 chip will not be available until August and will not be in blades. This makes the four wide a glueless config which has horrible performance since the Tuk chip only has 5 QPI's.

HP will only have a 16 socket 64 core system this year (August 2010). The p770 will best this machine easily.

The 32 socket system is June 2011....and the 64 socket system (four 8cellblade chassis's with connectors) does not have a date and if it does happen will be 2012.

Looks like you need a new NDA update.

Allison

0
0
Bronze badge

No it's about....

All the other stuff you need to do to put one in your data centre, and how energy inefficient they are.

BTW. It isn't anything to do with p series chips that reduces your licence costs for CPU based products. It is to do with how the OS domains / partitions CPU usage. Windows & VMWare aren't as good as AiX and Solaris at doing this. You can get the same sort of licence performance out of a Solaris box as you can from a P series because you can set hard limits on CPU usage.

Where P Series scores a bit better is on raw grunt over a Sparc, but that's not as much as it used to be with Sun's multi-threading chip architecture.

0
0
Megaphone

Domains/SPLPARS are hard partitioning

in Oracle terms.

So you can limit the number of processors you have to pay license cost for.

Now Oracle licenses on Niagara T2 processors is still 0.75 licenses per core. versus

1 license per POWER7 core.

And a 4 socket 750 is around 3 times faster than a 4 socket T5440, at equal number of cores.

So you still get x2 more umpf per license per core with power.

Furthermore you will run out of juice on a T5440 when you still have at least half the box left on a power 750.

Now the real trick with Oracle on power and sparc is that you can overcommit your ressources.

On power you can use the flexibility that the box offers you to really REALLY save on licenses. As you can make a processor pool and then run

An example. Lets say you have 4x 4 core power 520 servers that run Oracle 11. This will cost you 16 licenses. Lets say that they all vary in load and profile. Now what you do is you take and make a processor pool on eg a power 570 and then you move all your load onto this processor pool, as 4 seperate virtual machines. Before your move you have put your load figures together and sized this pool accordingly. You will most likely find that it only needs to hold 6-8 cores.

The result is that you get 4 virtual machine that run inside a processor pool of for example 6 cores. Hence.. you only have to pay for 6 oracle licenses. Now that is saving money.

On your T5XXX box or a MX000 , it is not quite as flexible, but you can make a domain or use the whole box and then use containers to do much the same thing.

But be sure to check that you don't fall in any pitfalls with regards to licenses, so that you don't end up having a Oracle sales rep that wants to charge you for the physical cores x number of containers/splpars. brrr...

And yes you can do it in one OS image without doing virtualization, but that will put serious restrictions on version etc.

// Jesper

0
0
Stop

The dilemma for IBM is...

I expect the Power7 top bin will be 10-25% faster than forthcoming Nehalem top bin speed.

But I also expect that Nehalem will be just 25% of the price.

And I also expect an x86 server will be 10% of the price of AIX tin.

0
0
Thumb Down

Your numbers are wront AC.

Well I don't know where you have those numbers from. I would say that top bin Nehalem-EX will be 10-15% slower than lowest bin (3.0GHz POWER7). But this is chip performance (specint), not system performance like for example a TPC-C, SAP 2/3 Tier.

If you look at Westmere versus Nehalem -EP then Intel had to raise the clock 13% to almost keep up with Nehalem-EP on the per core level. On a benchmark like specint_rate then the per core per GHz of Nehalem- EP is 19% faster compared to Westmere. Nehalem-EX won't be

You can see the same effect if you compare 4 core/chip POWER7 to 8core/chip POWER7. There is nothing surprising in this.

And I can only conclude that you don't really know what you are talking about, when it comes to price. You can't really buy a POWER7 chip, only the server.

So with regards to tin price. Then a current x3850 M2 with 4P and 128 GB ram costs around 40KUSD, a Nehalem-EX machine would most likely cost a little more.

A power 750 with 4P@3.0 GHz and 128 GB RAM costs 100KUSD.

But you also have to remember that the power 750 will be faster than a x3850.

If you factor this in, then I think that a factor of of 2 when it comes to price performance in specint kind of workload.

Currently the power 750 is way more than 2.5 times faster than current x3850. The current x3850 does 684,508 tpmc. We don't have any numbers for the power 750. But the old 16core@4.7GHz power6 power 570 does 1,616,162 tpmc with a rperf rating of 134.35. Now a 32 core power 750 has a rPerf rating of 292.47. Now I won't say that the power 750 will do 3.5M tpmc.

But if you compare current POWER7 to x86 gear using dunnington, I know where the price performance benefit is...

Sure a x3850 using Nehalem-EX would most likely do something between 1.5M-2M TPM-C.

So your 10x in price is.. well.. more likely somewhere in the factor of 1-2.

// Jesper

0
0

Jesper

"And a 4 socket 750 is around 3 times faster than a 4 socket T5440, at equal number of cores."... how have reached this conclusion? I mean, Sun Niagara shines at multi threaded work, I wonder how good the Power7 is at such work? As the 1.6GHz Niagara T2+ can have 5 times the performance of one 5GHz POWER6, I wonder if the POWER7 is 5 times faster than the POWER6? If POWER7 only is 3 times faster, then I find it hard to see how POWER7 can be "3 times faster" than a Niagara T2+ which can be 5 times faster than a POWER6?

0
0
Jobs Halo

Oh, let me repeat myself yet again

You simply cannot be bothered to look up the facts yourself now can you, and how many times does one have to repeat the obvious to you.

specint_rate2006

Power 750 -> 1070

T5440 -> 360

Difference is a factor of 3,0

documentation

http://www.spec.org/cpu2006/results/res2010q1/cpu2006-20100205-09566.html

http://www.spec.org/cpu2006/results/res2009q3/cpu2006-20090717-08200.html

specfp_rate2006

Power 750 -> 865

T5440 -> 230

Difference is a factor of 3,8

documentation

http://www.spec.org/cpu2006/results/res2008q4/cpu2006-20080929-05409.html

http://www.spec.org/cpu2006/results/res2010q1/cpu2006-20100205-09567.html

specjbb2005

specfp_rate2006

Power 750 -> 2478929

T5440 -> 841380

Difference is a factor of 3,0

Documentation

http://www.spec.org/osg/jbb2005/results/res2010q1/jbb2005-20100205-00800.html

http://www.spec.org/osg/jbb2005/results/res2009q3/jbb2005-20090720-00753.html

SAP 2 Tier

Power 750 -> 15600 users

T5440 -> 4720 users

Difference is a factor of 3,3

Documentation:

http://download.sap.com/download.epd?context=40E2D9D5E00EEF7C4759A696C959B04F207B1F9228664BDEE51955CADBEFB494

http://download.sap.com/download.epd?context=40E2D9D5E00EEF7CC663DDC83061BA6F9242BF63884801E0C4E97204702C1F27

So saying that it is a factor of 3 seems like being friendly to the T5440, as the average of the above is more like 3.3... but who is counting. So again.. do some research.. and get in touch with reality.

// Jesper

0
0
Silver badge
Happy

RE: Oh, let me repeat myself yet again

Jesper, whilst it is fun beating up on Kebabfart with benchmarks, it's a bit dishonest to pretend customers are guaranteed even three times the performance just going on vendor benchmarks. Sure, Power6 is much faster than Niagara in most cases in the real World, but reality does involve a lot more than just chip speeds. You also have to take into account the rest of your hardware (is it a memory intensive app rather than CPU intensive, is it heavy on disk access or LAN bandwidth, etc, etc), the app and OS stack, and the other systems it has to interact with. We had Power6 kit that trounced Niagara in one of our shoot-outs by a factor of two, which is lower than we expected, but was lower due to non-IBM parts of the overall solution so not IBM's fault. In day-to-day use we are actually seeing performance averaging about 10% less than what we saw in the shoot-out as it was very hard to mimic our real environment completely, and we didn't correctly predict an impact on one of the other parts of the solution. Again, not IBM's fault. So, for us, in this case Power6 in production turned out to be about 1.9 times faster than the best result Sun's reseller could get out of Niagara in the shoot-out. Still much better, and at a very good price too, so we're not too upset, but we might have been if we'd just assumed we were going to get 4+ times better performance as the original IBM reseller's pitch suggested.

Just to keep Kebabfart happy, I would like to point out we did have one shoot-out around a web-based project where Niagara outperfromed Xeon, Itanium and Power6. So, as they say, different horses for different courses.

If you already have Power6 in production, or have a very good idea that it will perform to the level you require, then this price-cut is great news! Power6 is proven and stable, and if you buy Power6 now you can let others find the bugs that are probably hiding away in the new Power7 and the new version of AIX, then upgrade later when IBM has fixes for them. Still, it will probably pay to pretend to the IBM salesgrunt that you are thinking of holding off until Power7 arrives to see if you can get a few more percent off the price.... ;)

/Shoot-outs - fun for everyone except the salegrunts!

0
1

I don't disagree with the essence Matt Bryant

First Matt all the benchmark results i mentioned in my post are POWER7. I haven't really said anything about POWER6, in this post.

Now according to Kebbubfert then Niagara is many times faster than POWER6, so your benchmarks gotta be wrong :)=

I have several times called Niagara for the best box in it's niche, with the only problem that it is to expensive compared to x86 iron.

Now as for benchmarks and test etc. Then it looks like you have spend to much time around sales people to listen to them. I've normally done my own calculations first, and then told them what I want. Not the other way around, it tends to be to expensive that way :)=

Also a big big factor in performance of a platform is the team that sets it up.

But IMHO Niagara is not as forgiving as POWER or Itanium, that both have good real life single threaded performance, hence when you hit a single threaded part of a solution it doesn't stall.

But You already know that :)=

And that can actually be something as trivial as starting an application up.

// Jesper

0
0

Jesper

I claim that on some work loads the Niagara is several times faster than POWER6, not in general. No one claims that, not even Sun. I would like to know how well the POWER7 fares in Siebel v8 benches. You need 14 POWER6 cpus to get half the benchmark score of four Niagaras in Siebel v8 according to official IBM and Sun benchmarks.

0
0
IT Angle

Some food for the drool

Usually you claim it's 28. I think I have debunked that one again and agin. Now this is the third of fourth time I repeat myself:

http://forums.theregister.co.uk/post/700377

// Jesper

0
0
Pint

Benchmarks vs Real World

I might disagree that SAP 2-tier is a system benchmark. It reeks of being trivial to run given the number of results and it never really showed how much NU is in everyone's NUMA. I've always believed more obvious NUMA systems like the Sun uniboard design with the slow interconnect (like the E25K boat anchor) had some way to keep offboard memory accesses minimal to non-existent. IBM Power seemed to be the only 3-tier show off.

Matt - I suspect your experience is like many others. Take TPC-C for example. Running a transactional database environment could be done a lot better in our shop (higher rates per unit of time) if I wanted to throw gaggles of 8Gb FC cards and more disk drive spindles and disk subsystems than I can afford at my configuration, not to mention memory. In an attempt to keep things 'fair' I've compared like configurations in my bake offs - say 4 FC paths and 4TB of disk in a single disk subsystem, because that's 'normal'. However the Power kit could probably drive more FC paths and more disks than the Sun kit. Who can afford a TPC-C kit - all those data paths and disk units and now flash storage...nice to look at, but no thanks!

I always liked TPC-C and SAP 3-tier benchmarks because they are really good full system benchmarks...and a system benchmark is always pretty easy to identify by the number of required network paths and disk paths configured to drive all the I/O.

0
0
This topic is closed for new posts.