[] IBM unveils Power8 and OpenPower pincer attack on Intel’s x86 server monopoly

[norseamd]

does it overclock?

can you run linux?

can you use it as a supercomputer?

 

[thunng8]

SAP benchmark results are pretty good.

Power8, 24 cores @ 3.52GHz: 115,870 SAPS

IvyBridge, 24 cores @ 2.7GHz: 51,650 SAPS

Also like thunng8 mentioned, I think the Power8's system specs should only list 2 processors, as the S824 only has two socket and the 12-core Power8 is clocked at 3.52GHz (the 6-core variant is clocked higher)

Yes, it does look overall POWER8 is approx double the performance of Ivy Bridge per core.

The s824 is a 2 socket system, but the benchmark results are correct. The s824 at least in the 24 core variant puts 2x6 core POWER8 chips into one socket called a DCM (dual chip module).

This was also done on some POWER7+ servers as well (the 750 and 760). My guess - yields are not yet good enough to sell a fully populated 12 core chip in volume. That may explain why the frequencies are slightly different from the OpenPower datasheets.

A Post on RWT on the POWER8 models available:
http://www.realworldtech.com/forum/?threadid=140955&curpostid=140955

Part 00NE368 - 12 cores - 3.425GHz base/3.891GHz turbo
Part 00NE369 - 12 cores - 3.126GHz base/3.625GHz turbo
Part 00NE370 - 10 cores - 3.425GHz base/3.891GHz turbo
Part 00NE371 - 8 cores - 3.758GHz base/4.123GHz turbo

 

[thunng8]

does it overclock?

can you run linux?

can you use it as a supercomputer?

1. It does have Turbo if that is what you mean.

2. Yes, RedHat Enterprise, Suse and Ubuntu

3. yes as well .. but there should be specific servers that are optimised for that later on (i.e. higher compute density).

 

[Lepton87]

does it overclock?

can you run linux?

can you use it as a supercomputer?

Of course it doesn't OC. Even if it was technically possible, no server board is going to offer any options to adjust clock-speed. You can't OC Intel newest Xeons either.

 

[DrMrLordX]

Not 100% sure but ORACLE/SUN cores were not licensable. The specs were so anyone can make a SPARC compatible core- but can anyone license the M6 core design and build their own CPUs with it?

Well, the T2 is available under the GNU General Public License. I'm guessing people could copy the T2 from whole cloth if they wanted, or just grab up parts of it. It does not appear that the T3, T4, or T5 are covered by the same license.

BTW, IBM has always almost been ahead of Intel in performance.. but that wasn't enough to lose market-share. They are now trying something different and making it licensable. i.e POWER7 was released in 2010. It took until late 2013 until the first Xeon cpu managed to edge past the POWER7 in performance (Xeon IVY Bridge EP and then EX). BTW .. the latest Ivy Bridge EX/EP only beats POWER7 since it has more cores per chip .. POWER7 still has the advantage per core on server workloads.

For their sake, let's hope that their decision works. Still, the simultaneous decision to open up the POWER8 design and dump fabs just looks . . . well, kinda bad really.

IBM have only announced the scale out servers. Enterprise servers will be coming later this year and scale up to 32 sockets.

Hmm. Well, that may be all well and good, but they aren't going for 4P scale out servers? I guess with the early performance numbers coming in, they feel they don't have to?

Should be interesting to see how POWER8 stacks up vs. Haswell-EX.

The openPOWER datasheet says 190W for the current models. (12 core 3.52Ghz or 8 core 4.1Ghz). Presumably the socket can take 250W though .. so maybe future version will be 250W but at higher frequencies.

If the author is to be believed, it's 4.5 ghz @ 250W TDP, vs. 3.52 ghz @ 190W TDP.

Of course it doesn't OC. Even if it was technically possible, no server board is going to offer any options to adjust clock-speed. You can't OC Intel newest Xeons either.

Hey now, you can overclock Opterons, even in 4p boards (for what it's worth).

 

[thunng8]

Hmm. Well, that may be all well and good, but they aren't going for 4P scale out servers? I guess with the early performance numbers coming in, they feel they don't have to?

Should be interesting to see how POWER8 stacks up vs. Haswell-EX.

No ideas of the exact model lineup will be .. but Power7 has 1 socket, 2 socket, 4 socket, 8 socket and a high end 32 socket server. Power8 should follow a similar path.

Haswell-EX won't be anywhere close to POWER8. POWER8 is 2X faster per core compared to Ivy Bridge. Haswell core is a known quantity .. approx 10% faster than Ivy Bridge.

 

[Fox5]

No ideas of the exact model lineup will be .. but Power7 has 1 socket, 2 socket, 4 socket, 8 socket and a high end 32 socket server. Power8 should follow a similar path.

Haswell-EX won't be anywhere close to POWER8. POWER8 is 2X faster per core compared to Ivy Bridge. Haswell core is a known quantity .. approx 10% faster than Ivy Bridge.

Haswell with code recompiled for it and AVX2 gets up to around 25% over Ivy. For the markets Power targets, people are going to be compiling software specifically for the cpu.
Still not enough to match Power8, but performance per watt is what really matters here.

 

[DrMrLordX]

Still not enough to match Power8, but performance per watt is what really matters here.

That's what I was getting at. There seems to be some kind of disconnect between what the article says about the perf/watt situation between POWER8 and some unspecified product of Intel's (presumably Ivy Bridge-EX) and what others are saying about the same matchup.

If POWER8 is twice as fast as Ivy Bridge-EX and uses 190W vs. 155W of Ivy Bridge-EX, then the perf/watt is heavily tilted in favor of POWER8. Yet, the article author (who is quite the cheerleader for IBM) claims that Intel still has the perf/watt crown.

And, all the while, the actual competition that POWER8 will be facing throughout its market lifespan will be Haswell-EX (or a later descendant), not Ivy Bridge-EX. Haswell-EX is going to be at least 10% faster than its predecessor at a lower TDP.

So, that is why I think direct benchmarks between Haswell-EX and POWER8 should prove to be instructive. Some of the available data is a bit hazy at this time. It may just be that the article author has no idea what he/she/it is talking about here.

 

[Ill_take_Power]

Before I respond with my thoughts on why I disagree with the perf/watt claims as having no relevance I would like to hear why you think they are. From my perspective they are relevant only when placing into a form factor and it's ability to cool and survive the environment. If not, who cares? Sure, if you have 10K x86 servers you may, but with Power you can put those 10K on 100 servers. It isn't about perf/watt but perf/core. Ah oh, I couldn't hold off on my thoughts. Seriously though, help me understand why I should care about perf/watt. Thank you

 

[Phynaz]

Seriously though, help me understand why I should care about perf/watt. Thank you

You aren't responsible for paying the electric bill for a data center, are you?

 

[DominionSeraph]

Not sure how you can compare a complete system to a single component (CPU).

I was comparing systems. You can pull 12 core systems off Dell Outlet for $2k.

 

[rtsurfer]

You aren't responsible for paying the electric bill for a data center, are you?

Even if the Power CPUs consume more power, they would finish the task faster than Intel CPU & race to Idle faster.

It should even out in the end.

 

[NTMBK]

The statistic that really matters is not performance/watt, but performance/joule.

 

[Ill_take_Power]

No, I don't pay the power bill but I happily work with data center managers to show them how server consolidation dramatically reduces power as well as rack space, pdu ports, network ports, san ports, air handlers/chillers, UPS, etc.

This is why I asked why performance/watt matters relative to performance/core. Yes, it matters (I don't want to dismiss it because somebody would say what about 1KW or higher) but when the core is stronger, can do more work, drive utilization levels, host more VM's reliability and securely then a processor having a 25 - 50% higher TDP sounds pretty reasonable if it can reduce the footprint by 20X, yes? Further, the TCO of the solution is greater than just the infrastructure savings described above but also the software, maintenance and even FTE associated with it.

Power8 vs Ivy Bridge EX, Haswell XYZ, SPARC T 12345678 and whatever is the same. IBM delivers strong performance per core. The virtualization is secure, flexible and scalable supporting Unix, Linux (LE & BE) and IBM i.

 

[TuxDave]

The statistic that really matters is not performance/watt, but performance/joule.

If...

Performance = instructions / sec
Watt = joules / sec

Then...

Performance / Watt = instructions / joule and that describes the total amount of work completed per unit of energy. However, I haven't quite figured out what instruction / (sec * joule) would describe so maybe I'm missing something.

 

[Ajay]

hmm... temporarily lost my mind

 

[Phynaz]

Even if the Power CPUs consume more power, they would finish the task faster than Intel CPU & race to Idle faster.

It should even out in the end.

Huh? He asked why performance per watt matters. What you just described is performance per watt.

 

[Ajay]

Still not enough to match Power8, but performance per watt is what really matters here.

Yes, from the little I know about datacenters, it's really about performance/$. Over a 3 year period - the cost of electricity for power and cooling become dominate and the single largest consumer of power and heat production is the CPU. A step up from that is very high availability servers. Really advanced RAS features, usually including the ability to 'hot swap' the OS core and components for patches with minimal down time. I have no idea how big that market is or where it's going.

So Power8 looks pretty good against IVY-EP/X, but the next inflection point for Intel comes with Broadwell-EP/X. Haswell-EP/X will offer mainly an improvement in FP performance and maximum core count (which may have up to 18 cores). I don't think the core count goes up with Broadwell, but the power consumption goes way down.

I imagine Power8 will span at least one die shrink (to 14nm FinFet) like Power7 did (45 & 32nm).

 

[thunng8]

Some more details of third parties starting to test their Power8 motherboards:

http://www.enterprisetech.com/2014/04/28/inside-google-tyan-power8-server-boards/

Tyan has an AT-X system board and Google has already designed (made by Foxconn) and are internally testing.

Looks like Tyan one shown is a lower end 1 socket board and the Google one has 2 CPU sockets.

 

[DrMrLordX]

No, I don't pay the power bill but I happily work with data center managers to show them how server consolidation dramatically reduces power as well as rack space, pdu ports, network ports, san ports, air handlers/chillers, UPS, etc.

This is why I asked why performance/watt matters relative to performance/core. Yes, it matters (I don't want to dismiss it because somebody would say what about 1KW or higher) but when the core is stronger, can do more work, drive utilization levels, host more VM's reliability and securely then a processor having a 25 - 50% higher TDP sounds pretty reasonable if it can reduce the footprint by 20X, yes? Further, the TCO of the solution is greater than just the infrastructure savings described above but also the software, maintenance and even FTE associated with it.

Nobody has demonstrated how or why choosing POWER8 over, say, Ivy Bridge-EX amounts to consolidation. Today, I can choose an "off the shelf" 4P Xeon board and match it up against a 2P POWER8 board. According to the article author, I can expect both systems to deliver about the same performance (varying some based on the workload) with the Xeon system consuming less power. The Xeon system lets me get away with a weaker PSU, less localized cooling, less stress on my air conditioning, and less power drawn from the wall.

If I ignore the article and instead look at the TDP from data sheets and then look at SAP results, I can assume that the 2P POWER8 solution will deliver the same performance as the 4P Ivy Bridge-EX solution @ ~61% of the power consumption, making POWER8 dominant in perf/watt.

Either way, choosing the solution with the best perf/watt lets you spend less money on operation and infrastructure. So, is it POWER8 or Ivy Bridge-EX that has the best perf/watt? You tell me. I would prefer not to make inferences based on the dismissive comments of POWER8 advocates.

Even if the Power CPUs consume more power, they would finish the task faster than Intel CPU & race to Idle faster.

It should even out in the end.

So long as you choose the solution with the best performance per watt, and so long as your solution is fully scalable, you can just keep adding nodes until you get the performance you want. What stops you from throwing more CPU/cores at a workload? Power consumption. How much performance can you fit into a fixed power budget? That's what server buyers face, and that's why perf/watt matters. The solution that races to idle fastest is the one that can deliver the best performance on the available power.

If POWER8 were some 10+ ghz monstrosity known for clock speed and IPC, then maybe it could race to idle faster on lightly-threaded workloads, but that is not the case. POWER8's mo is massive thread parallelism per socket.

Power8 vs Ivy Bridge EX, Haswell XYZ, SPARC T 12345678 and whatever is the same. IBM delivers strong performance per core. The virtualization is secure, flexible and scalable supporting Unix, Linux (LE & BE) and IBM i.

I highly doubt that you'll find SPARC T5, Ivy Bridge-EX, and Haswell-EX to be the same.

What?! Watts and Joules are both measures of power; might as well suggest using BTUs or horsepower would be better....

Joules are a measurement of energy. Watts is joules/second, a measurement of power. Not exactly the same thing.

Yes, from the little I know about datacenters, it's really about performance/$. Over a 3 year period - the cost of electricity for power and cooling become dominate and the single largest consumer of power and heat production is the CPU.

We have a winner.

So Power8 looks pretty good against IVY-EP/X, but the next inflection point for Intel comes with Broadwell-EP/X. Haswell-EP/X will offer mainly an improvement in FP performance and maximum core count (which may have up to 18 cores). I don't think the core count goes up with Broadwell, but the power consumption goes way down.

It's taken Intel a loooong time to bring Haswell to the server room as their flagship processor. In fact, they haven't quite done that yet (June/July I think?). Broadwell-EX will be awhile yet, unless they aren't serious about letting Haswell-EX sit on top of the Xeon product line for very long.

If Haswell vs Ivy Bridge in the desktop is any indicator of how Haswell-EX will perform, you'll see improvements from AVX2, and you'll see nice improvements in TDP at the same clock speed, but the clock speeds won't increase by much (if at all).

 

[tsupersonic]

Damn, really? We are just talking about getting some new P7+ servers at work to replace our P7's. Too bad we can't wait for P8

 

[Ajay]

Joules are a measurement of energy. Watts is joules/second, a measurement of power. Not exactly the same thing.

I completely mis-wrote that comment. I should have my BS in Physics revoked :$

It's taken Intel a loooong time to bring Haswell to the server room as their flagship processor. In fact, they haven't quite done that yet (June/July I think?). Broadwell-EX will be awhile yet, unless they aren't serious about letting Haswell-EX sit on top of the Xeon product line for very long.

If Haswell vs Ivy Bridge in the desktop is any indicator of how Haswell-EX will perform, you'll see improvements from AVX2, and you'll see nice improvements in TDP at the same clock speed, but the clock speeds won't increase by much (if at all).

I think that Broadwell will make a larger impact. The advantage Haswell brings over IVY won't be as big as Broadwell will bring over Haswell - mainly a 30% decrease in TDP at a given clock rate, which will lead to more significant savings in power bills. And, BW still has AVX2 and all the architectural improvements that HW has.

 

[thunng8]

Some comparisons of Ivy Bridge and POWER8:

SpecInt_rate_2006
2.7Ghz Intel Xeon E5-2697 v2 - 24 cores - 934
3.52Ghz POWER8 - 24 Cores - 1750

Specfp_rate_2006
2.7Ghz Intel Xeon E5-2697 v2 - 24 cores - 649
3.52Ghz POWER8 - 24 Cores - 1370

Source:
http://spec.org/cpu2006/results/res2014q1/cpu2006-20140121-28204.html
http://spec.org/cpu2006/results/res2014q1/cpu2006-20140121-28203.html
http://benchmarkingblog.wordpress.com/2014/04/28/awesome-power8-benchmarks-awesome-dessert/

Another interesting article with leaked Intel performance estimates for future Haswell and Broadwell performance on the above benchmarks:

http://wccftech.com/intel-xeon-proc...-2014-broadwell-ep-18-cores-45-mb-cache-2015/

i.e. Broadwell 18 core will still not match Today's 12 core POWER8 .

If we extrapolate from the graphs:

SpecInt_rate_2006
Top bin E5-2697 v4 Broadwell - 36 cores - ~1450
3.52Ghz POWER8 - 24 Cores - 1750

Specfp_rate_2006
Top bin E5-2697 v4 Broadwell - 36 cores - ~1000
3.52Ghz POWER8 - 24 Cores - 1370

POWER8 is not standing still in the meantime either with higher clocked POWER8 chips to be released as well as future POWER8+ shrink.

 

[Ill_take_Power]

Well stated @thunng8! What @DrMrLordX chooses to overlook is you don't compare a single x86 server to a single Power server. Most businesses with 1000 employee's and higher will have a an IT staff, maybe an operations staff and 1 or two data centers of varying sizes and sophistication. I'll accept we can run a Power server (it can be P7 or P8) at 90% utilization - let's just say 100% of TDP. The server could be a 1, 2, 4, 8 or 32 socket server depending on the environment. Because of the efficiency of the Power Hypervisor it is doing the work that a 12, 16, 24, 32, ... core x86 or SPARC server can do and do it with 1/4th or 1/8th or 1/20th (I am just picking ratio's I have used in the past against these platforms) the compute resources. Now, run Oracle on that x86 server at 25% util on a 24 core server (which is 6 effective cores) it will require 12 licenses @ $47,500/license. That is roughly $600k + 22% maint per year starting with year 1. On P8, let's say only 3 cores are required (it's my story so I can tell it how I want ) that would be 3 Oracle licenses at $150K + 22% maint per year. Most x86 shops will deploy more x86 servers for each Oracle workload. For Power, they will just stack them on the same server. If we add a 2nd workload the x86 is another $600k totaling $1.2M + 22% maint whereas the Power is $300k + 22% maint. See how this scales? I don't want you to think I'm calling your baby ugly for no reason. The reality is, x86 vendors are positioning their servers to run these enterprise workloads where Power, SPARC, PA-RISC, Itanium, Alpha, MIPS and others have been for decades. Performance per core is key. When you have to buy server after server after server like x86 then perf/watt makes perfect sense. Just don't try to apply what is important to you to Power as it isn't relevant.

 

[NTMBK]

If...

Performance = instructions / sec
Watt = joules / sec

Then...

Performance / Watt = instructions / joule and that describes the total amount of work completed per unit of energy. However, I haven't quite figured out what instruction / (sec * joule) would describe so maybe I'm missing something.

Heh, totally correct, I didn't work through my units fully. I should hand back my BSc Useful computation/J is indeed the one I was thinking of... which is equivalent to performance/W. Doh!