IBM Releases Power7 Processor

Dan Jones writes "As discussed here last year, IBM has made good on its promise to release the Power7 processor (and servers) in the first half of 2010. The Power7 processor adds more cores and improved multithreading capabilities to boost the performance of servers requiring high up-time, according to Big Blue. Power7 chips will run between 3.0GHz and 4.14GHz and will come with four, six, or eight cores. The chips are being made using the 45-nm process technology. New Power7 servers (up to 64 cores for now) are said to deliver twice the performance of older Power6 systems, but are four times more energy efficient. Power7 servers will run AIX and Linux." And reader shmG notes Intel's release of a new Itanium server processor after two years of delays. The Power7 specs would seem to put the new Intel chip in the shade.

Comments

4.14GHz?

[Anonymous Coward]

What happened to the "3GHz ceiling"? Why can IBM go above it but Intel, AMD and VIA are stuck below it?

Re:4.14GHz?

[MichaelSmith]

Different architecture maybe? A little bit like how RISC could clock faster than CISC back in the day.

Re:4.14GHz?

[Totenglocke]

You mean how you can buy a 3.4 GHz Phenom II X4 from AMD? That 3.0 GHz ceiling?

Re:4.14GHz?

[the linux geek]

Its still that way. POWER6 could actually go up to 5GHz, but IBM sacrificed out-of-order execution to get there. POWER7 brings it back with a slightly lower clock speed and more cores.

Re:4.14GHz?

[wizardforce]

There probably were better ways of increasing computational speed using multicore processor designs than just increasing the clock speed. Kind of like going from a V4 engine to V6 being a better option in terms of power than increasing the individual piston HP of the V4 from 25 to 30.

Re:4.14GHz?

[alvinrod]

Actually both AMD and Intel have chips currently clocked over 3 GHz. Some of the newer Intel chips also have something they call Turbo Boost where the chip essentially overclocks itself if it's not using all of its cores. It also looks like AMD has a 3.6 GHz Phenom II X4 chip slated to be released soon. It would appear that the companies found solutions to whatever ceilings may have existed. VIA doesn't target the high-end of the market so I don't think that they're producing any chips that would run at those clock speeds.

Re:4.14GHz?

[RightSaidFred99]

First, there is no 3GHz ceiling, so you're begging the question. Second, these processors use specialized cooling - not run of the mill cheapo barely-enough heatsinks. If AMD or Intel spent $20 more on their heatsinks, they'd easily be selling 3.4-3.8GHz processors. But the profit margin isn't there. Third, power usage hikes as you increase voltage high enough to hit those speeds. Most people running nuclear explosion simulations on a 4GHz processor don't care, people running 30,000 machines in a design center...do care.

Re:4.14GHz?

[Hurricane78]

If AMD or Intel spent $20 more on their heatsinks, they'd easily be selling 3.4-3.8GHz processors. ...
Third, power usage hikes as you increase voltage high enough to hit those speeds.

You’re contradicting yourself. The reason they can in fact not easily ramp up the CPU speed, is exactly this increase in voltage. Which increases temperature at a cubic speed relative to processor speed. (See the Pentium 4, for what that results in.)

And because your bring not a single actual argument to why you think there is no 3GHz ceiling (actually it’s a gray area above 3 GHz), I call your argument... busted! ;)

Re:4.14GHz?

[RightSaidFred99]

There's no contradiction. Power usage is less of an issue on higher end "enthusiast" chips. They could easily sell 3.6GHz chips in this space with better heatsinks (as evidenced by...people running them at 4GHz easily on air cooling).

In the commodity space, even with better cooling, the power usage increases disproportionately as voltage goes up. There is a sweet spot, and it isn't currently >3GHz.

Finally, I didn't point out why there is no 3GHz ceiling because it takes 30 seconds of googling to see that there are currently chips selling at > 3GHz, and there have in the past been x86 CPUs up to 3.73GHz.

Busted my ass.

Re:4.14GHz?

[Anonymous Coward]

Don't forget that one of those 4GHz CPUs probably costs over 10 times as much as an equivalent Intel or AMD part. A decent PC costs as much as a car payment -- a decent POWER machine costs as much as a car.

The price is old, but a couple years ago a 5GHz Power6 CPU cost $15k for a dual-core module (with 4 threads) plus $30k to activate each core. That means you'd pay $75k total to use both cores of the CPU module. I'm sure Intel would have no problem supplying 5GHz CPUs at $75k each, but it's unlikely that they'd have many takers, so you're stuck with CPUs that are only 3GHz (but go almost as fast as IBM's 5GHz parts).

dom

Re:4.14GHz?

[wizardforce]

Cost. Spending an extra 500$ to double the power makes sense. Spending 5,000$ to increase the power a measly 20% is rather foolish either way you look at it by comparison.

Re:4.14GHz?

[afidel]

Intel did 3.73 Ghz as the top end for Netburst (Xeon 5080) but it was a fairly poor performer on a MIPS/Watt basis. In fact the 5160 running at 3.0Ghz did about 33% higher Specfp and run at 80W instead of 130W for the 5080 (35.2 specfp_2000/watt vs 15.5).

Re:4.14GHz?

[TubeSteak]

There probably were better ways of increasing computational speed using multicore processor designs than just increasing the clock speed. Kind of like going from a V4 engine to V6 being a better option in terms of power than increasing the individual piston HP of the V4 from 25 to 30.

Back in my day, manufacturers used to slap a turbo button on the front of the case.
And we liked it that way.
Now get off my lawn!

Re:4.14GHz?

[amorsen]

Spending 5,000$ to increase the power a measly 20% is rather foolish either way you look at it by comparison.

Not if your work load doesn't scale with additional cores. Then $5000 for 20% extra speed can be worth it.

Re:4.14GHz?

[hairyfeet]

The reason why AMD and Intel don't push the high end chips is frankly there just isn't much of a market for them ATM. Since most of the games being released are for console first and PC second, they simply aren't being bottle-necked by the CPU. This is also why AMD and Nvidia are having to push multiscreen and GPGPU, because frankly a less than $100 card will play a good 80%+ of the games out there.

Second for the jobs the average Joes are doing, web surfing, music/video, maybe the occasional video conversion, even the lower end chips are well past "good enough" for them. I have been selling a lot of low end AMD dual and triple core machines lately, and all I hear from my customers is "how fast" they are, and how they never seem to slow down. With hardware acceleration on the motherboard these 2.4GHz-2.8GHz duals and triples are frankly overkill, with most of the time the CPU twiddling its thumbs. I myself bought a 925 quad when they got so cheap, but a good 90% of the time the chip is barely above idle.

So it isn't that AMD and Intel can't make them, because we have seen in the past they can, it is just there really isn't much of a market for them. To get faster than 3.2GHz you really start cranking up the heat and the power, and that equals higher electric bills most folks don't want, not to mention having fans that sound like a F15 taking off isn't very pleasant. With the new 95w like my 925 the chips rarely get above 83f idle and so far mine has maxed at 109f doing video transcoding. And the 65w duals are so quiet I have to watch when setting them up I don't turn them off when I mean to turn them on, because i simply don't hear any noise. Folks nowadays seem to care much more about that than the MHz race anymore.

Re:4.14GHz?

[Your.Master]

You learn something new every day.

Re:4.14GHz?

[couchslug]

"Back in my day, manufacturers used to slap a turbo button on the front of the case.
And we liked it that way."

Noobs...
Back in MY day, I used to wax the strings on my abacus to lessen bead friction.
We LOVED it that way.
Now get off my peat bog!

Direct comparisons are bad

[the linux geek]

POWER and Itanium are architecturally so different that kdawson's snide "put this new Intel chip in the shade" comment is kind of nonsensical. Itanium is superscalar to an extent that POWER doesn't come close to, with each core being able to execute up to six instructions per cycle. While its possible that POWER7 is faster, its also more expensive to get a reasonable configuration and the performance difference between the two is not as clear-cut as our illustrious editor is trying to suggest.

Re:Direct comparisons are bad

[Anonymous Coward]

The Itanium is more that just superscalar, it is explicit parallelism. You can accomplish the same feat with superscalar and out-of-order execution but it takes far more silicon and it tends to have some odd corner cases.

POWER and Itanium are both pretty slick architectures but Itanium is definitely a generation later in design. If only Intel were willing to bet the company on it, about 10 years ago, we would all be using it today.

Re:Direct comparisons are bad

[Wesley Felter]

POWER and Itanium are architecturally so different...

That doesn't matter; they both address the same market (high-end Unix) and thus they are competitors.

Itanium is superscalar to an extent that POWER doesn't come close to, with each core being able to execute up to six instructions per cycle.

Yeah, POWER7 can only execute... six instructions per cycle. And you might indeed say that an in-order Itanium at 1.7 GHz doesn't come close to an out-of-order POWER7 at 3+ GHz.

While its possible that POWER7 is faster, its also more expensive to get a reasonable configuration...

Since no Tukwila servers have been announced, we don't even know how much they will cost.

Re:Direct comparisons are bad

[the linux geek]

http://www.tpc.org/tpcc/results/tpcc_perf_results.asp - I wouldn't say "clobber," but they're roughly at par on performance and Itanium has an edge on price/performance.

Re:Direct comparisons are bad

[MrNemesis]

Since no Tukwila servers have been announced, we don't even know how much they will cost.

As a sysadmin for a company with POWER5 and 6 equipment, all I can say is if you have to ask you can't afford it. Part of the reason why jumping ship to RHEL + Oracle running on a VMware cluster is looking increasingly appealing to managment.

Re:Direct comparisons are bad

[xZgf6xHx2uhoAj9D]

That leaves you with emulation, which i doubt Intel could make faster than native...

If only you could go back in time and convince Intel of this! The first generation of Itaniums actually did x86 emulation in hardware. A brilliant idea: the only problem with it was that it was actually slower than software emulators (which themselves were pretty slow).

Anyway I don't think Itanium was every supposed to replace x86. This was before x86-64 existed and Intel thought it would be their only 64-bit chip.

Intel can't move to a new architecture because they are held back by all the millions of closed source applications out there.

Ahh but they did! It's called x86-64!

In the end it wasn't backwards compatibility that was the problem. x86-64 has the almost all the same backwards compatibility problems that Itanium has: software developers are forced to release two binaries for their code these days, an x86-32 binary and an x86-64 binary. Obviously x86-64 took off like gang-busters, though, even though it came out like 5 years after Itanium. The reason developers are happy to develop x86-64 binaries but not Itanium binaries is that there still doesn't exist a decent compiler for Itanium.

Okay, okay, I know x86-64 trivially does x86-32 "emulation" very efficiently, which helped it out, but I think the existence of extremely good compilers is what helped it more.

Re:Direct comparisons are bad

[imgod2u]

That whole philosophy went out the window when Intel couldn't make a compiler good enough to make Itanium work well in all situation; which to this day -- despite having more software engineers than silicon guys -- they still don't.

Scheduling things beforehand will only get you so much. It sounds good on paper in a "look how much silicon we save" kind of way but the reality is, explicit parallelism and static scheduling simply aren't good in this day of variable memory latencies, multi-tiered caches and people no longer programming in assembly.

Real question

[LordoftheChmod]

That's all fun but it doesn't answer the real question : Can it run Crysis?

Re:Real question

[skine]

I think you clicked the wrong link on Tom's Hardware.

The question here is whether it can run Linux - followed shortly by a debate on how terrible Ubuntu is.

Re:This is Bad News

[RightSaidFred99]

Right, because the Power series is such a global moneymaking powerhouse, right?

These chips (and Itanium) are niche. In fact, the commodity chips are getting so powerful across the entire CPU segment from embedded to HPC computing that they will start eating into the market of even these niche chips. Why buy a Power7 when you can buy 3-4 Nehalems, be twice as fast, and spend 1/2 the money? There are exceptions, but they're becoming fewer and fewer.

Ah, AIX

[wandazulu]

AIX....the last Unix you can't just "get" a copy of, but need to actually buy the hardware (a la the Mac). We had a Power box at work with AIX for awhile, but its configuration tools was quite ... unique among Unix flavors (though I was told it was pretty straightforward IBM) and I had a horrible time getting GCC to work with it; most every F/OSS package I came across either straight up wasn't tested on AIX (because no one had the hardware), or it had a whole separate setup (I believe one of the standard lines running ./configure is "Is this an AIX system?").

I recall the box being wicked fast when we were running Oracle on it; it was a "small" Power machine but it still could handle a monster database with hundreds of millions of rows with no trouble. Frankly, I was sort-of sad to see it go; I really did want to get more familiar with it, but apparently the maintenance costs IBM was charging made it a non-starter. Plus, ultimately, it seems that it just wasn't very OSS friendly; xlc is apparently an amazing compiler for the PowerPC, but they wanted $6000 for a license per developer. Plus, and I'm sorry if this is nitpicking, but to have the C compiler called xlc and the C++ compiler called xlC was just, well, insane.

What I really wanted to do was get Linux on it, and Oracle even has a Linux-on-Power version of their database, but there seemed to be some grumbling from the IBM salespeople (according to my boss) that they discourage people from running Linux on Power....I guess you (according to them) need AIX to unleash the real "power" in the PowerPC.

Sigh, okay, whatever. back to Linux on x86-64.

Re:Ah, AIX

[argent]

AIX....the last Unix you can't just "get" a copy of, but need to actually buy the hardware (a la the Mac).

Don't forget HPUX.

Re:Ah, AIX

[Junior J. Junior III]

And Mac OS X.

Apple skunkworks?

[bobdotorg]

I'm curious whether or not Apple is maintaining a parallel dev. of OSX for this line of IBM chips the same way that the Intel version of OSX was lurking in the dark from 2000 until 2006.

Re:Apple skunkworks?

[Khyber]

LOL You're in no position to know.

I am, however. But my NDA forbids me from saying anything.

Re:Apple skunkworks?

[Khyber]

Actually, I just checked said NDA, as it has been at least five years since I worked for them, and my NDA is over, so....

Support has NEVER been fully dropped and never will unless IBM becomes non-viable in the marketplace. On top of that, long-term contracts Apple has with some companies pretty much ensures that they keep some minimal amount of POWER support active, at least into the next decade.

Oh, I'm sorry, did I break your bullshit reality bubble? Get a real job in the industry and maybe you'd have half a clue.

Commercial sales?

[CAIMLAS]

I have to wonder why IBM is (at least, as of now) limiting these processors in their own hardware.

I can understand the initial economic advantage: they'd gain more profit from server sales, and would be able to sell Linux servers at a fairly non-trivial mark-up (on base hardware cost, to them).

But what is gained there is probably trivial compared to commercial marketing of the chips/boards (OEM sales). I suspect it might also avoid scrutiny from antitrust lawyers more easily. Why wouldn't they do this? I'd certainly love a processor like that; it'd be incredible. 1/4th the power envelope of the Power6, and twice the performance (assuming it means core clock)? That's incredible: the 3.2GHz Power6 is rated at under 100W TDP.

Such a processor might just sway Apple to go back to the Power architecture, I'd think. Linux will run on them, obviously; the only thing you couldn't run on them is Windows (and even that might be possible down the road with only a little work on MS's part).

The only two reasons I can imagine are 'exclusivity' and 'insufficient fab capacity'. That second one would certainly do it on its own.

Re:Commercial sales?

[Wesley Felter]

A non-IBM POWER7 system would end up looking pretty much like an IBM POWER7 system, and you can bet it wouldn't be cheaper, so what's the point? If you want POWER7, buy it from IBM.

Re:Commercial sales?

[Anonymous Coward]

I don't know what a 3.2GHz Power6 costs, but last I checked a 4.2GHz Power6 cost $12k! Somehow I don't think Apple will be swayed to using them unless IBM can sell them at 1% of their current price.

dom

Re:Commercial sales?

[rtaylor]

You don't spend $10,000 per CPU then put in a small amount of crappy ram and a single tiny SATA harddisk.

Uh, did you look at your link?

[raftpeople]

Power procs=32, cores=64
Itanium procs=64, cores=128

So double the Itaniums almost gets you to where Power is.

Re:Uh, did you look at your link?

[RightSaidFred99]

What's your point? All that matters is price/performance. If processor A can get the job done more cheaply with twice the processors, why would I care?

Re:Uh, did you look at your link?

[raftpeople]

Because this statement was made: "but they're roughly at par on performance", which isn't correct.

Your point is valid, but that isn't what I was responding to.

Re:Uh, did you look at your link?

[Tycho]

I'd also add that depending on the task, the cheap solution would be slower if the task had serial parts that could not be separated into threads. For instance if a task takes 1,000 cycles and all of the instructions must be done in a precise order, a quad-core processor running at 2.0 GHz would be slower and be of lower utility than a single core 4.0 GHz processor, assuming all other things are equal. The quad-core ends up working at half the speed of the single core and the quad-core also has the penalty of three idle cores draining electricity.

I would also imagine that these newer POWER7 processors carry over the decimal floating point units present in the POWER6. Yes, floating point units that operate in base-10 as opposed to base-2. Not necessarily of much value for scientific purposes, but great for preserving accuracy in financial calculations. One gets to avoid the base-10 to base-2 conversion and the conversion back that can severely hurt accuracy with only a binary floating point unit. One also gets a nice speed up by doing decimal math in hardware as opposed to the other option of software decimal math.

Re:Uh, did you look at your link?

[Calinous]

You pay some licensing costs PER PROCESSOR (or PER CORE, or a combination thereof). As such, it might be cheaper to buy a $100k, 32 processors server than a similar performance, $50k, 64 processors server.

Re:Uh, did you look at your link?

[Bert64]

Because performance per watt is also important, if you have twice as many physical processors and twice as many cores you will also need all the supporting infrastructure (ie sockets for those processors to go in) etc... Not to mention the extra space required.

And more power consumption also equals more cost.

Otherwise, why use powerful machines at all, why not a cluster of the cheapest machines you can find?

Incidentally, something which performs half as well needs to be considerably cheaper or it just won't sell at all.

Re:Uh, did you look at your link?

[DarkOx]

Price / Performance is not all that matters. If it was we would probably be running tens of di-shrunk 486 cores in PCs today. Complexity matters as well. Nobody wants to deal with hardware nonsense in software. In the PC world you can't even get people to write threaded apps; and you're going to tell me its ok to ask developers to deal with 2x as many cores to get the same amount of computing done? That is going to complicate you application a great deal!

From your link

[raftpeople]

"The PowerPC 970 is derived from POWER4. It lacks some server oriented features, but does have an AltiVec unit. The 970 and its descendants are used by Apple and IBM and some high end embedded applications."

LPARs

[Gothmolly]

IBM gear gets you LPARs, with a real hypervisor that is laps ahead of all the other stuff.

Re:LPARs

[nelvinboy]

IBM gear gets you LPARs, with a real hypervisor that is laps ahead of all the other stuff.

You're absolutely right. We're just finishing migrating our data warehouse Database, ETL, and BI systems from smaller x86 boxes over to LPARs on a mid-upper-range POWER6 box. The performance and the on-the-fly configuration that our AIX admin can do on this box is SIMPLY AMAZING. We have lots of boxes running in VMWare environment, too. But the capabilities there performance-wise don't even touch IBM's visualization.

Re:Query

[Wesley Felter]

Anyone have data on how these compare to x86 and Intel's latest creations? Presumably, one could write an efficient algorithm for a variety of common computing tasks and port it to the different chips to get a cross-architecture performance estimate.

That's called SPEC CPU; here are some results: http://www.realworldtech.com/forums/index.cfm?action=detail&id=107244&threadid=107238&roomid=2

Re:Query

[scotch]

Holy shit, this guy is on to something. You could write these common computing tasks as a sort of "bench" suite of tests. Then on each architecture, you would get different "marks" against the "bench". Let's call them "benchmarks" for brevity. These "benchmarks" would give allow clear and unambiguous comparison of these various chips. Foolproof and brilliant!

Re:Query

[afidel]

On a $/FLOP basis they get slaughtered by Nehalem-EX, but if you need flat out performance the Power7 system will be superior thanks to 2x more memory bandwidth per core and ~3.5x more interprocessor bandwidth. The basics for this type of comparison are Specfp_base, Specint_base for CPU performance and usually either SAP, TPC-C or specjbb for business logic comparisons.

Re:Query

[Jeremy Erwin]

I like your folk etymology. It neatly excises surveying from the discussion.

Re:This is Bad News

[Anonymous Coward]

You'd buy a Power7 because it comes with 63 other Power7 friends in a single box and runs an operating system specifically designed for the ridiculous number of cores and capable of handling even the most data intensive legacy applications.

I agree that the high end server market is becoming smaller and smaller as time goes on but in reality there's still a huge backbone of legacy applications that require the sort of processing throughput only a single whopping great server can provide. The kind of applications that draw $150,000 3 month contracts for developers because nobody knows a damn thing about them, general public included.

CPU speed vs memory bandwidth, I don't get it

[egnop]

I can understand why you would get a Power chip for pure number crunching.
But having a lot of data to chew away, I use p-threading for the larger jobs and let the rest of the jobs over to the os.

I was always under the assumption that data has to be delivered to the cpu fast, very fast and since the Power6 rs6000 only supports ddr2 I don't get it.

We recently bought a new rs6000, which has the Power6 in it(still has to be delivered), but the memory is 'only' ddr2, can someone enlighten me why this machine would run faster than my dual Xeon 5560 with triple channel ddr3?

The Xeon box only costed 1/2 of what the rs6000 costed

thanks in advance

Re:This is Bad News

[Anonymous Coward]

You buy it for the hypervisor, massive IO, and capacity upgrade on demand. Forget what you know about virtualization from xen and VMWare. The POWER hypervisor lets you add (or remove) ram, buses, and processors from a running server. You can even set the memory and cpu to pull from a shared pool (with set priorities and limits). The internal 10Gb network doesn't hurt either.

Try scaling your xen system when you are IO bound to disk. POWER offers physical and (fast) virtual IO, giving each partition "big iron" IO capacity.

POWER is made by people who understand scaling. Commodity boxes are made for people who like big numbers printed on the side of the box and don't understand why high CPU and memory numbers are useless if your disk array can't keep up.