IBM lifts the veil on Power7 chips

With every new server processor generation, there is always the possibility that the chip and its associated chipsets and system components are so different that both the chip and the server platform have to change at the same time. This will not be the case for the Power7 processors due next year, says IBM. IBM is in no hurry …

[Anonymous Coward]

Curious how IBM announces future Power7 on day that Sun announces faster CMT

IBM's Power7 "futures" announcement today is just smoke and mirrors to offset Sun's announcement which was made today about the launching/shipping of faster UltraSPARCT2 plus 1.6GHz systems -which by the way is already more than twice as fast as IBM's current Power6+ @ 5GHz!

http://www.sun.com/aboutsun/pr/2009-07/sunflash.20090721.2.xml

[Anonymous Coward]

Huh...

2-3x perf is obviously accomplished with more cores/threads... not going to 15 Ghz... or even 6 GHz...if P7 is just a multi-thread chip then you already get that with T2+ from Sun... as comment above P6 performance * 2 = Sun T2+ at 4X the cost... I guess this means that single thread performance from IBM will start decreasing now and single threaded apps on 4.7 GHz P6 may actually run slower after upgrade to P7 3.? GHz.... awesome.

[Anonymous Coward]

Serial Number only??

I especially like this:

Customers "...will be able to upgrade to Power7 machines in such a way that they can preserve their serial numbers on their boxes."

Should I presume that a serial number is the only thing which will be preserved? :-)

[Anonymous Coward]

@AC Serial Number only??

>> Should I presume that a serial number is the only thing which will be preserved? :-)

That's pretty much how they did in-box upgrades from p5 to p6... you basically kept the skins and the main power switch which had the serial number plate attached to it... slightly different from just swapping out CPUs...

it's pretty clear that all your memory will also need to be swapped out (DDR2 to DDR3) - that's usually an expensive component in these sorts of system

of course still no indication of doing anythiong like Sun DSDs or HP nPars... so one failure can still take out your entire 595 and all those 1000 VMs you run on it... eggs.... baskets... etc.

[ToddRundgren]

IBM Power 7

UltraSPARC 2-3 X faster than Power 7. I find that very very hard to believe. What are you measuring?

[Anonymous Coward]

@ AC - About Sun T2

Don't compare IBM's P6 or P7 with Sun's T2. They are complete different architectures.

For example, if you need more than 4 sockets on T2 where do you go with no Rock processor in the future to scale?

IBM's architecture scale from 1-socket blade to 32-sockets High-end boxes. Sun has to prove their CMT architecture can do the same for a fair comparation.

It's very easy to boost performance in 1 to 4 sockets server (e.g. x86). The real challenge is to go beyond and Sun's CMT is not even close to do that.

It would be nice to see a comparison between Sun's CMT versus Intel Nehalem and check if Sun has real advantages on the low-end segment chips up to 4 sockets.

Cheers!

[Anonymous Coward]

The new Binary Compatibility

...will be able to upgrade to Power7 machines in such a way that they can preserve their serial numbers...

so this is the binary compatibility IBM refers to?

Come on!!!!! Get us something better than this!!!

[Anonymous Coward]

Sounds like Power7 is a wonderful story

Let's not bash IBM to defend SPARC gear who's future is now dependent on Larry's patience.

Let's all agree to give Power7 a chance to be tremendous technology vs. be defensive. It's time to cash out and move on.

Jonathan

[Anonymous Coward]

T2 faster than Power6? Only if you solder it to F16...

"UltraSPARCT2 plus 1.6GHz systems -which by the way is already more than twice as fast as IBM's current Power6+ @ 5GHz!"

That is interesting what you say. Could you compare 32-core to 32-core, please?

The website you pointed to says sometihing different. In SAP benchmark, 32 SparcT2 cores are 26% faster than 8 Power6 cores. That means single T2 core is 3.2 times SLOWER than single Power6 core. Please do the calculations yourself.

[Anonymous Coward]

@ AC about @ AC - About Sun T2

"For example, if you need more than 4 sockets on T2 where do you go with no Rock processor in the future to scale?"

uh... SPARC64... Scales up to 64 sockets... Complete binary compatibility.

It's all SPARC with different price points.

[Anonymous Coward]

T2 actually beating Power

SPECint_rate2006

*

The Sun SPARC Enterprise T5440 server, equipped with four 1.6 GHz UltraSPARC T2 Plus processor chips, delivered 57% and 37% better results than the best 4-chip IBM POWER6+ based systems on the SPEC CPU2006 integer throughput metrics.

*

The Sun SPARC Enterprise T5240 server equipped with two 1.6 GHz UltraSPARC T2 Plus processor chips, produced 68% and 48% better results than the best 2-chip IBM POWER6+ based systems on the SPEC CPU2006 integer throughput metrics.

*

The single-chip 1.6 GHz UltraSPARC T2 processor-based Sun CMT servers produced 59% to 68% better results than the best single-chip IBM POWER6 based systems on the SPEC CPU2006 integer throughput metrics.

*

read more here: http://blogs.sun.com/BestPerf/entry/1_6_ghz_spec_cpu2006

[David W Johnson]

T2 beating Power .. waiting to see results

Not that I don't believe vendors blogs because they are always right, but it is interesting that spec.org site hasn't posted SUN's latest SPECint_rate2006 results. I can't find any new results for the T5440 or the T5240 servers on the actual spec site (sorry, not going to read companies A blog against company B). The only spec info regarding SUN in the month of Jul-09 is the Sun Blade 6048. So are they waiting to post this new info on the spec website?

The only thing I see is the old spec result. I like to compare apples to apples if possible since benchmarks stress the cpu, memory and compiler. So looking at the old results, not sure who would be better. The current info on spec.org website for the Sun T5540 result had twice as much memory than the Power 550 box so if the 550 box had the same amount of memory, would it be faster, slower, I'm not sure. Same thing goes for the T5240, twice as much memory than the Power 520 (even though I would argue why you would buy a low end box like the T5240 or 520 when you could get a Intel Nehalem box from HP, IBM or even SUN for less money).

Does it really matter now? SUN systems road map is cloudy at best so most UNIX shops will be migrating from either HP or IBM. It was nice to have a 3 UNIX vendors competing for customers dollars, but now the sun has faded for SUN.

[Jan 7]

Which single failure takes a p595 offline?

Whoa, loads of AC flogging the dead remains of SUN to bash IBM.

You know its a weird day when you have one -ever so tiny- sympathy for Matt B.

anyway: Could the AC from 1652 (4th post) please tell which single failure is able to take down a p595? Or a 570, for that matter? I´d be greatly interested.

[Anonymous Coward]

T2 actually beating Power??

T2 is an 8-core CPU, Power6 is 2-core CPU. You have to compare core to core. Processes run on core, not on socket.

[Kebabbert]

SUN Niagara way faster than IBM. Proof:

Anonymous Coward,

"That is interesting what you say. Could you compare 32-core to 32-core, please?

The website you pointed to says sometihing different. In SAP benchmark, 32 SparcT2 cores are 26% faster than 8 Power6 cores. That means single T2 core is 3.2 times SLOWER than single Power6 core. Please do the calculations yourself."

Well, maybe you havent understand the statement, but I will try to explain to you. He states that one 1.6GHz Niagara CPU is faster than a 5GHz Power6 CPU. He is not discussing core vs core. He is discussing CPU vs CPU. Because you dont buy 1-2 cores. You buy a whole CPU. Right?

Now let us see... If I state that one Niagara is faster than a Power6, does it mean that the Niagaras ALU unit is faster? Or the integer register is faster? Or the Bandwidth is faster? Or one core? Or does it mean that the CPU is faster? What do you think? Could you explain what you think if I say that "this CPU is faster than the other CPU"?

That link you discuss, shows that 4 of the Niagara 1.4GHz CPU is 26% faster than 8 Power6 5GHz CPU. Now, which CPU is faster? Niagara or Power6? (Hint: you can not buy one core, you buy a whole CPU. It is meaningless and dumb to compare parts of a CPU, when we discuss the CPU as a whole)

Here we see 3 of the IBM Power P570 servers with 12 Power6 CPUs at 4.7GHz benchmarked in SIEBEL against 1 SUN T5440 with 4 Niagara CPUs at 1.4GHz. That single Sun machine is twice as fast as 3 of the P570 servers. The SUN machine achieves 14.000 SIEBEL, and 3 of the IBM Power P570 achieves together 7.000 SIEBEL. Further down here, you see links to Oracle web page with white papers with benchmarks, proving this:

http://blogs.sun.com/mandalika/entry/siebel_8_0_on_sun

Here we see price comparison

http://www.sun.com/servers/coolthreads/t5440/

IBM P570 with 4 of the 4.7GHz Power6 and 64 GB RAM costs 413.000 USD.

SUN T5440 with 4 of the 1.4GHz Niagara and 64 GB RAM costs 76.000 USD.

You get a fraction of the performance with an IBM Power570 and it costs many times more than one Sun T5440

Anonmous Coward

"T2 is an 8-core CPU, Power6 is 2-core CPU. You have to compare core to core. Processes run on core, not on socket."

You know, you buy one CPU and you the whole of the CPU. You compare Niagara CPU against IBM CPU, right? Because the statement is "Niagara CPU is faster than Power CPU". The statement was not "Part of the Niagara CPU is faster than Power CPU". So let me ask you, which CPU is faster? Now I talk about the CPU. Which processor is fastest? Can you answer that?

[ShaggyDoggy]

Serial numbers

The reason that the serial number is important to iSeries (AS/400) is that it plays a part in s/ware licencing, device naming, hardware config, etc etc.

Changing QSRLNBR is not done lightly

[Anonymous Coward]

@AC about @ AC about @ AC - About Sun T2

"uh... SPARC64... Scales up to 64 sockets... Complete binary compatibility.

It's all SPARC with different price points."

uh... I was talking about CMT processors.

But hey, If you are talking about midrange and high-end gear, POWER has been kicking Fujitsu's SPARC64 butt for a long time...

I was talking about Sun gear, Sun only has to low-end gear to offer now. Fujitsu is a different story.

One might recall that Sun would replace all SPARC64 gear this year to the dead Rock.

Yes, you have to agree, if a customer wants more than 4 chips T2+ they have to buy Fujitsu stuff, that's sad.

Regards, MM

[Anonymous Coward]

@ Jan7

" anyway: Could the AC from 1652 (4th post) please tell which single failure is able to take down a p595? Or a 570, for that matter? I´d be greatly interested."

I am not the AC from 4th post.

I just want to say to Sun zombies lovers that maybe they don't know that both POWER 570 and 595 have concurrent maintenance for sometime.

That means, you can repair live machines without powering it off.

But hey, IBM also offers partition mobility with Live Partition Mobility , Concurrent Firmware updates , Concurrent AIX Updates , and a long list of RAS features.

Cheers!

[Kebabbert]

Anonymous Coward

Anonymous Coward,

"I was talking about Sun gear, Sun only has to low-end gear to offer now."

But you know, only because SUN boxes has low prices, it doesnt mean they are low-gear. One SUN T5440 is 80.000USD. And one IBM P570 is 413.000USD. And we know that one T5440 with 4 of the 1.4GHz Niagara CPUs, is twice as fast as three of the IBM Powerserver P570 with 12 of the Power6 at 5GHz, on SIEBEL benchmarks? It is funny that Sun's "low-end gear" handily outperforms three p570. Maybe SUN should multiply their prices with 10, then you wouldn't consider SUNs boxes as "low-gear"?

Europes largest ISP, STRATA, which handles 1 billion email/day migrated their whole backend to one T5440. Another company, migrated 251 Dell dual Core Linux servers, running 700 instances of MySQL, down to 24 of these T5440 boxes. And notice, this is the "low-gear" capable of this stuff. The Niagara uses the least power on the entire market on heavier chips, ca 100 Watt it is because of the low 1.4GHz frequency. Maybe Power6 at 5GHz uses 500watt? If you are not able to get performance with low frequency, then you have to reach 5-6GHz, and that will not do. Because a normal 2-3GHz CPU will idle 50% of the time according to Intel, during full load because of cache misses. 50% of the time during full load, a cpu has to wait for cache data. The faster the CPU, the more it has to wait. One cpu at 5GHz has to wait 70% of the time because of cache misses?

The new planned SUN T3 Niagara will have 16 cores, each having 16 threads. That is 256 threads. Roughly one of these chips corresponds to one T5440. Rumours say that SUN is planning a box with 8 of these T3 CPUs. That is 2048 threads in one box. Luckily Solaris scales extremely well, so it can handle all these threads. That will yield extreme throughput.

On the other end, Fujitsu's new Sparc64 CPU codename Venus, will have 8 core cores. It has 128 GFlops and is the fastest CPU on earth. Google on "Venus" to read more on this beast. Rumours say that SUN ROCK cpu would not be faster than the Fujitsu monster Venus. That is why SUN canceled ROCK. Why develop and sell Rock and Venus at the same time?

Soon you can run Enterprise Solaris with it's unique and revolutionizing feature set on massive throughput Niagara T3 servers or on massive number crunching Sparc64 servers. For a fraction of the price. And all the competition will, as usual, bite the dust.

[Anonymous Coward]

@Kebabbert

Dude, again don't tell us that a single T5440 beats three P570's for god's sake.

This is ONE benchmark and it fits into the T2 CMT's architecture nicely. Good for Sun...

Maybe you don't have experienced large corporate datacenters systems and applications.

It's obvious that emails, web servers, application servers will fit nicely on CMT servers.

But, I doubt that a large corporation like banks, retailers and such will buy T5440's to run their SAP or Oracle EBS.

Core apps just WON'T fit on these low-end gear, wake up...

About T3, is it another Sun vapor? Who knows? But, if it scales up to 8 sockets it will be nice for Oracle.

About Fujitsu Venus, when it will be deployed? By the time it is on the streets it will have to prove it can match the POWER7 performance.

Nor Sun or Fujitsu can talk about roadmaps, people just don't buy it anymore.

Cheers!

[Anonymous Coward]

@ac @AC about @ AC about @ AC - About Sun T2 #

"Yes, you have to agree, if a customer wants more than 4 chips T2+ they have to buy Fujitsu stuff, that's sad."

So? What's your point. This is why Sun opensourced SPARC. It's not the first time that Sun has used someone elses tech. IBM resells NetAPP's storage. HP resells Itanium... Everyone resells Intel/AMD... Who cares. The fact is that you can run SPARC all the way at the low end and all the way up to the high-end, and all running Solaris with binary compatibility.

At the same price point IBM is nowhere near what CMT can do. IBM has to go to Intel to get anything near CMT. I think that is sad. That all said, I agree you should not compare a T5440 to a high-end 570 as the 570 is aimed at the highly available, vertically scaling environment.

[Anonymous Coward]

@ac @ac @AC about @ AC about @ AC - About Sun T2 # #

"The fact is that you can run SPARC all the way at the low end and all the way up to the high-end, and all running Solaris with binary compatibility."

Dude, I am not IBMer I am just a customer who uses lots of POWER boxes and reads the news.

However, Sun did promise to have High-end gear of its own. After Sun canceled the Rock chip, going OEM for higher end boxes, that's very sad in my opinion.

My point is Sun's T2 CMT don't scale and don't fit for all types of workloads.

For that price range, you are right. IBM can't compete using dual core POWER chips on low-end against Sun T2, x86 quad, hexa cores. That's why I told you guys to compare T2 and Nehalem.

But, if one needs neat virtualization, thread performance, POWER is still unbeatable. That's my opinion.

Cheers!

[Anonymous Coward]

@Kebabbert # By Anonymous Coward Posted Wednesday 22nd July 2009 15:34 GMT

>This is ONE benchmark and it fits into the T2 CMT's architecture nicely. Good for Sun...

>Maybe you don't have experienced large corporate datacenters systems and applications.

>It's obvious that emails, web servers, application servers will fit nicely on CMT servers.

>But, I doubt that a large corporation like banks, retailers and such will buy T5440's to run their SAP or Oracle EBS.

Sun T5440 World Record SAP-SD 4-Processor Two-tier SAP ERP 6.0 EP 4 (Unicode)

Sun SPARC Enterprise T5440 Server World Record Four Processor performance result on Two-tier SAP ERP 6.0 Enhancement Pack 4 (Unicode) Standard Sales and Distribution (SD) Benchmark

* World Record performance result with four processors on the two-tier SAP ERP 6.0 enhancement pack 4 (unicode) standard sales and distribution (SD) benchmark as of July 21, 2009.

* The Sun SPARC Enterprise T5440 Server with four 1.6GHz UltraSPARC-T2 Plus processors (32 cores, 256 threads)achieved 4,720 SAP SD Benchmark users running SAP ERP application release 6.0 enhancement pack 4 benchmark with unicode software, using Oracle10g database and Solaris 10 OS.

* Sun SPARC Enterprise T5440 Server with four 1.6GHz UltraSPARC T2 Plus processors beats IBM System 550 by 26% using Oracle10g and Solaris 10 even though they both use the same number of processors.

* Sun SPARC Enterprise T5440 Server with four 1.6GHz UltraSPARC T2 Plus processors beats HP ProLiant DL585 G6 using Oracle10g and Solaris 10 with the same number of processors.

http://blogs.sun.com/BestPerf/entry/sun_t5440_world_record_sap

[Anonymous Coward]

For those that still don't understand CMT

Why does 1.6 beat 4.7?

Sun has upgraded the UltraSPARC T2 and UltraSPARC T2 Plus processors to 1.6 GHz. As described in some detail in yesterday's post, new results show SPEC CPU2006 performance improvements vs. previous systems that often exceed the clock speed improvement. The scaling can be attributed to both memory system improvements and software improvements, such as the Sun Studio 12 Update 1 compiler.

A MHz improvement within a product line is often useful. If yesterday's chip runs at speed n and today's at n*1.12 then, intuitively, sure, I'll take today's.

Comparing MHz across product lines is often counter-intuitive. Consider that Sun's new systems provide:

* up to 68% more throughput than the 4.7 GHz POWER6+ [1], and

* up to 3x the throughput of the Itanium 9150N [2].

The comparisons are particularly striking when one takes into account the cache size advantage for both the POWER6+ and the Itanium 9150N, and the MHz advantage for the POWER6+:

http://blogs.sun.com/BestPerf/entry/why_is_1_6_ghz

[Anonymous Coward]

'That means, you can repair live machines without powering it off.'

I nearly wet myself.

I can't recall the last time I did any system maintenance at a major UK bank that didn't require an outage on rs6000/system P/pSeries (still sounds like 'pessary' to me). pSeries stuff is so far behind 'mainframe' class systems for in-situ repairs. You still can't replace memory or cpu while the box is up which is such a PITA since the parts IBM sources fail constantly.

Oh you meant disk swaps? Yeah, you can do those without shutting an rs6000 down... usually.

[Anonymous Coward]

@AC RS/6000 misinformed dude

Dude, RS/6000 brand is gone not yesterday but 8-10 years ago.

After that, it was pSeries and then POWER servers.

Please, update yourself before posting.

Try starting reading, this:

http://www-03.ibm.com/systems/power/hardware/whitepapers/595_570_cec.html

That is related to concurrent maintenance of high-end servers that ONLY 'mainframe' class systems are enabled to..

Best regards! Cheers!

[Anonymous Coward]

@AC 'Please, update yourself before posting. '

Same shit, different badge - you did notice I said 'rs6000/system P/pSeries' didn't you?

I use p570's on a daily basis, and I can say that popping ram chips, MCM's and RIO books happens so frequently that I'd love it if I could hotswap with the system on line. Yeah I know the new flexy interconnect helps coldswapping a node out, but with a virtualised load which is overcommitted as per IBM's own best practices please explain how I remove 4/8 cores from a system already running at capacity?

I followed your link in case I missed something from last weeks IBM sales propaganda sessions... I guess not...

ftp://ftp.software.ibm.com/common/ssi/sa/wh/n/pow03023usen/POW03023USEN.PDF

Not Found

The requested URL /common/ssi/sa/wh/n/pow03023usen/POW03023USEN.PDF was not found on this server.

IBM_HTTP_Server/6.0.2.13 Apache/2.0.47 (Unix) Server at ftp.software.ibm.com Port 80

[Anonymous Coward]

@AC Not Found

The link worked for me, and just tried again and it works

" IBM Power 595 and 570 Servers

CEC Concurrent Maintenance

Technical Overview

April 14, 2009 "

ftp://ftp.software.ibm.com/common/ssi/sa/wh/n/pol03021usen/POL03021USEN.PDF

"I can say that popping ram chips, MCM's and RIO books happens so frequently that I'd love it if I could hotswap with the system on line"

Something must be wrong with your environment, we have a dozen of p570s here untouched for months and a couple with 1 year+ of uptime.

"overcommitted as per IBM's own best practices please explain how I remove 4/8 cores from a system already running at capacity?"

Another IBM best practice in these cases is to have Live Partition Mobility-ready servers, so you can move partitions easily without disrupting your apps.

Best regards! Cheers!

[Kebabbert]

Anonymous Coward

"Though, I have a feeling you're exaggerating quite a bit here since a "major UK bank" experiencing the frequent outages you describe would switch to something else ASAP."

Oh, there are politics involved. The best tech doesnt necessarily wins. How many uses sucky Windows, and never switch to Linux? In my company, large fortune 500, Ive recommended new technology that would make enormous savings and better up time. But to no avail. The management has decided on some technology and that's it. I can do nothing, despite showing better calculations.