It has been more than a year since IBM got its first Power7-based machines out the door, and about six months since the chips were fully ramped across the Power Systems lineup. The server processor racket waits for no one, and a slowpoke will quickly get left behind in the volume and midrange space. And so Big Blue has to …
The problem IBM is facing is that it can't spread the R&D costs for a new Power chip over the same volume of sales as Intel is able to for the x86 architecture.
The obvious solution would have been to buy Apple, and license the 680x0 architecture from Freescale, and put together PowerPC Macintoshes that shipped with chips that seamlessly ran both PowerPC and 68020/68882 code at native speed, and which shipped with both OS X and Power PC and 680x0 editions of OS 9, all running concurrently.
Either in a dual-boot configuration that also includes a PowerPC port of OS/2, or with the base operating system allowing either the OS X user interface (with OS 9 present) or Presentation Manager to be started from an OS/2 text prompt.
Including, of course, quality keyboards made by Unicomp.
And, unlike the Macintosh, the hardware platform would be open, like the IBM PC. So, when other manufacturers make peripherals for the new standard in computing, the choice of peripherals and software would be initially comparable to, and hopefully shortly greater than, that for the obsolete Windows/Intel platform that IBM abandoned.
Of course, that's a nice trick if you can pull it off. The IBM PC did not succeed because of the IBM name on it - it succeeded because it was the first 16-bit system that wasn't much more expensive than a serious business 8-bit CP/M system. The IBM name helped to give reason to believe that software would be as available for it as for CP/M (and, hey, in one standardized disk format)... but that's all.
There is no way that is obvious to me that IBM could today make a desktop PowerPC-based system that was so obviously superior to a typical Wintel PC as to persuade people to switch over. This is unfortunate, because that is the trick that IBM needs to pull off. Otherwise, the PowerPC could end up forgotten before the Itanium.
What are you banging on about? This has very little to do with the article, which is in no way talking about desktop PCs. These are chips for midrange servers. And it won't be "forgotten before the Itanium" because Power chips have already been around for 20+ years and a lot of corporate customers rely on these. Their use is growing pretty well, especially in the Unix world with Sun in turmoil as Oracle prefers unbreakable Linux and HP-UX being years behind.
Now, if IBM wanted to get back into the desktop PC market (which they clearly don't) they could put some R&D into a Power version of Linux. Not sure who would want this though.
Second all that.
Plus not to forget the success of PowerPC in the almost invisible embedded market where Freescale have been earning good money in the the telecommunications sector for their PowerQUICCs. And then there is the good old 8641D which, despite running at a miserable 1.3GHz, is still quicker at some sorts of useful maths than Intel chips.
powerpc linux is surprisingly good
I wanted to see the recent progress and get ready for the future end of os x support so I downloaded the gentoo powerpc distro.
My last experience with Linux on G5 was horrible as the machine decided to run all the legendary 14 fans at full speed.
Even on a hardcore advanced linux, gentoo, which was traditionally optimised for x86 perf. , the fans didn't go crazy, nvidia was performing at lcd native resolution and there was amazing speed difference on things like building gcc. Please note it is the rarest quad g5 (ppc970mp) from Apple. On a fully documented IBM powerpc blade etc, things would be even better.
So not to confuse soon be abandoned powerpc mac owners, I thought I better post it.
What IBM needs is, finding things that is exclusive to x86 (valgrind?) and make them function on POWER4+ command set. Also contribute (anonymously?) to projects like ffmpeg which are becoming/already became standard in huge media transcoding facilities which are perfect match for their power blades.
Completely off topic...
Mmm, I'd like to see three or four Power7 cores in the next Wii, but maybe four Power6+ cores on 45nm is more likely given the target die size. And most likely is a simple reboot of the XBox/Cell PPU ... bah :-(
And still going off topic even further...
I hear from distant contacts in the games industry that they're just beginning to get to grips with the PS3's CELL processor (for it is a mightly complicated beast) and discovering with joy the raw power that's available to them. I fear that Sony might just ditch the CELL processor, and that would be a great pity.
The problem they've had is that to program a CELL well you need a background in high speed real time multi-processing applications. The systems that get used for building modern military radars and the like are architecturally very similar to the CELL processor, and indeed the CELL processor has found some uses in those fields. The problem has been that not many people in the games industry have come from those highly specialised application areas, so the games industry has had to re-learn some of their tricks.
Seems back to normal then
Given its only until 2001 that the roadmap spiked up and now seems to be going back to a more patient line of upgrades. I'm sure the marketing types will have less sleepless nights having to push kit out the door in a year before its outdated.
All that said AIX seems to have been ramping up in versioning in the past decade as well, still wished they were to release a educational x86 version in some form, my AIX version 4 sertification seems a waste nowadays.
I would also point out that the roadmap isnt complete with the power .1 chip aka the RT6150 box's which had the RT(Risc Technology) chip inside, they were realy after all were the birth of the power chip lays back in the days were you had AT(Advanced Technology) XT(eXtended Technology) based computer systems or PC's as some would say.
What I find more interesting in the later and later generations of power kit is not the CPU love but more the supporting chipsets that bring more and more of the mainframe resiliance down to the cheaper kit.
Love them or hate IBM there aint many original chip manufacturers that have carried there torch's as long. Though the days of a sysadmin are somewhat being moved towards a split of VM handler and VM designer, still fun kit if you like to get yoru geek on. Even if AIX is UNIX with a registery :o).
"Even if AIX is UNIX with a registery :o)."
True, my only gripe with what is an otherwise fine OS, the ODM doesnt go wrong often and you can fix it on the go, but when you have to wrestle with several DB's which all contain the same shit just to clear out a bad software install then it gets very very tedious..
Give me my path_to_inst and flat files any day of the week.
Least yoru not having to deal with older versions of MQ series were you would find a simple OS upgrade on what would appear to be transient spaces would stamp all over your MQ queue's as they were located (ug cant recall but somewere silly).
Still nomatter what I lub AIX, always to me seemed to have the nice balance of BSD and AT&T flavours of old school NIX and with that lovely JFS filesystem and partitioning control, even in the early days was streets ahead on that front.
Yeah I can imagine the fun on a bad ODM from a software install fail, though its the hardware changes were you don't notice anything and the box is in a HACMP cluster and later down the line you find corrept deviced entries/dupes and other quirks, those are the real fun times. But I will always swear by a mksysb, even if I only want to backup the ODM, me motto being backup don't fudgeup.
Still long past are the days of installing the OS of around 80 5 1/4" floppies only to find disc 67 corrupt and to repeat the whole process the next day using your resellers copies to find there disc #67 is also corrupt. Or the fun of seeing a manager do a rm the OS and you dive across hitting all forms of break and finding out there is no copy of the OS to reinstall as the kits so new in the country and you end up getting creative.
But its a lovely operating system and I'd take a running man over a hourglass anyday of the week ;0.
Well, what it had to do with the article was this: IBM is at a disadvantage because it doesn't sell lots and lots of POWER chips the way Intel sells lots and lots of x86 chips.
Which is why they're lagging behind with new chip generations, the way the article notes.
Getting back on the desktop is, of course, wildly impractical - but that seems to be what IBM would have to do to fix this problem. Unless there's a better idea, will IBM have to give up, leaving ARM as the only alternative to x86?
That's not how IBM operate
Your still not getting the point. IBM don't really sell POWER chips. They don't really sell computers either.
What IBM do sell is apparently quite expensive business services (which does include some hardware), and looking at their profitability you'd have to say that they're clearly value for money. IBM's silicon needs merely reflect their need to keep selling business services. If they can do that with what some might argue is old fashioned out of date silicon processes and chip designs then IBM will be quite happy with that. Indeed, busting a gut to build a faster chip that doesn't help sell more business services would be commercially idiotic.
Developing their own POWER chips does allow IBM to tailor the silicon to the needs of the business services that they sell. For example, ever wondered why the POWER chips have a decimal maths FPU as well as a standard FPU? Why would IBM go to all that effort when no one else ever has?
It's because for banking / financial applications standard double precision FPUs on Intel/AMD chips are not accurate enough, so you have to do the maths a different way. For example, doing $$$billions currency conversions need to come out to the last snippet of a cent, and double precision binary floating point maths doesn't get you that.
On an Intel chip you have to do that decimal maths in software (a bit like the bad old days of having an 8086 without the 8087...). It's slow and time consuming. But a POWER processor does the decimal maths for you, so it ends up being much quicker than an Intel x64. Which means for certain banking applications IBM can offer a service that's much quicker / cheaper / power efficient than someone offering a solution based on Intel processors. And so far that 'niche' market is big enough for IBM to make very impressive profits indeed.
Basically there is a level of sophistication to IBM's business model that escapes most people's attention, and is completely different to Intel's. You just can't read anything of consequence into comparisons of IBM and Intel hardware. You can compare IBM's and HP's business service offerings, and I'd say that HP (who just happen to base theirs on x64/Itanium) don't appear to be as good. If IBM ever decide that they're better off with Intel or anything else, they'd drop POWER quicker than you can blink.
So that's IBM's cool headed, commercially realistic side. Then they go crazy and do things like the IBM PC (a long time ago, but still crazy considering what IBM's core business was), or the CELL processor. The CELL processor in particular promised the world tremendous compute power, got used in the PS3 and had the high performance embedded market buzzing with anticipation.
And then they drop it just like that because it turns out that most of the rest of the world was taking far too long to learn how to use what is unarguably the most complicated and 'different' chip that anyone has produced in recent years, so they weren't selling enough to make it worth their while. Grrrrrr! Pity in a way - I think that if they had persisted then they would have cleaned up eventually because there's aspects of the CELL which are far superior to the GPUs that have started filling the void.
IBM are a tremendous technology company, but don't often give the niche markets the things that that they could. That's capitalism for you!
I dont agree with you much, bazza.
"...It's because for banking / financial applications standard double precision FPUs on Intel/AMD chips are not accurate enough, so you have to do the maths a different way..."
So, that is the reason POWER has a decimal math and standard FPU? Well, you are wrong. Financial and banking applications never do any rounding, they never use an FPU. Everything is computed as integers, and the number of decimals are book keeped separately. So, when you do a computation, you specify the amount of money, and you specify the nr of decimals. This way you dont never do any rounding. If IBM POWER are doing rounding, they are amateurs having no clue about finance nor banking.
Regarding why IBM stopped manufacturing the "superior" CELL cpu, you are wrong again. It is not because people could not appreciate nor program the CELL tech, as you claim. That is just IBM marketing talk. The reason CELL stopped, was because it didnt live up to expectations. It was bad, quite simply.
For instance, in String Pattern Matching, a heavy optimized CELL version (loop unrolling, assembler, etc) done by IBM researchers, bit the dust when compared to a SPARC Niagara T2+. The T2+ running at 1.4GHz is 13x times faster than a 3.2 GHz CELL. The T2+ just used a plain C version of the algorithm, copied from an algorithm book with no optimizations at all. It is pretty bad when a 3.2GHz using heavy optimizations by researchers can not beat a 1.4GHz Niagara, dont you think? You need more than 10 CELL cpus totaling 32GHz of aggregate Hz, to match one Niagara which has 1.4GHz. The CELL construction is very ineffective. One Hz from Niagara is as fast as 23 Hz from CELL. The ratio is 1:23 - which is not very effective at all.
It is just IBM that tells everybody that CELL is superior, but people dont have the wits to appreciate nor program it. The truth is, the CELL sucked quite bad. Otherwise it would not be discontinued.
Really?! You're going to quote a guy who puts this at the top of his page:
Thoughts from an unofficial Anonymous Sun employee"
Now that just screams objective, non-biased reading.
"So, that is the reason POWER has a decimal math and standard FPU? Well, you are wrong. Financial and banking applications never do any rounding, they never use an FPU. Everything is computed as integers, and the number of decimals are book keeped separately. So, when you do a computation, you specify the amount of money, and you specify the nr of decimals. This way you dont never do any rounding. If IBM POWER are doing rounding, they are amateurs having no clue about finance nor banking."
You're missing the point again. Decimal floating point by it's very nature avoids rounding.
We've suffered with binary floating point for a long time because it can be implemented with far fewer transistors and is good enough most of the time. Now that we have billions of transistors per chip, decimal floating point hardware is now feasible.
Modern banking application use decimal floating point types. The recent standard is IEEE 754-2008. Lots of languages have support for this. Java has BigDecimal. When an application that uses BigDecimal is run on p or z series machines with an IBM JVM, it uses the hardware instead of doing everything in code.
Keb'. You are simply wrong.
Well, I work in finance. A large world famous company. And I KNOW we dont use floating numbers. We use integers and avoid rounding. I suggest people to start to use integers instead of decimal floating numbers. To round is amateurish at best, and wrong.
This is my discipline. I work here. You are wrong.
So what are you trying to say? What is your point? That the benchmark is made up and never done? It is a lie? Well, it is easy for you to implement the same well known algorithm and run it and see. I never doubt that the numbers were actually achieved, because that would mark the source as a liar, and no one wants that. It is too easy to repeat the same benchmark and disprove the benchmark.
Sure we don't know how the actual code looks at your company. But having written 'banking code' myself years ago. I know that the code I wrote, and which most likely is still running down in germany somewhere then it heavily depended on software libraries doing DFP.
And honestly I think you have misunderstood something. Your statement 'We use integers and avoid rounding.' more suggests that you are using DFP rather than a Integers.
Try to read:
Another completely impartial piece from TPM....
"..... Sun Microsystems (now part of Oracle) started slipping ..... Fujitsu did the same ..... Intel started flubbing ...... Meanwhile, IBM took its foot off the Moore's Law gas...." So, if there's any slippage it's only with manufacturers with not called "IBM", but if a company called "IBM" have a slip it's "taking its foot off the gas"? What a laugh!
At least we now know why TPM was launching attacks on the coming top-end Tukzilla kit, it's because there's an IBM launch coming up!
Matt is getting testy
Maybe because everyone is shocked to hear that the 32 socket box is still not shipping. When I was at Oracle world I stopped over at the HP booth to compare technology as I always do. I was told last October the 32 socket was on the web shipping and the 64 socket is an easy extension with two more cables.
Apparently its a "32 Socket Starter Package" which is a little deceptive if you ask me.
Even the brochures say its there with a little disclaimer *3 32 Socket Starter Package; upgrades to 32 socket server in future release
HP would be wise to be low key about the release because they will alert everyone that they are about 8+ months late.
Obviously there will not be a 64 socket expansion as the brochure would have had 64.
So what HP is limping around with is a 16 socket box that has the same number of cores as IBM's 8 socket box with each core only being about 40% of Power7 core (compare the 1TB results at tpc.org). So today's full SD2 is about 3 sockets in a 4 socket p750.
RE: Matt is getting testy
Actually, Matt was laughing! Something I also often do in response to your posts, Allipoos.
As I understand it, the 32-socket Starter Package allows you to have 32-sockets in one bundle, managed and deployable as one lump, but not able to use the whole 32-sockets in one hardware partition as of yet. So you have to split it into two 16-socket hardware partitions but have control of both partitions as though they were one server. For consolidation, where you usually lumping lots of smaller (IBM or SPARC) servers onto hp hardware partitions or Integrity Virtual Machines instances, that's not a problem. So it is a 32-socket system, despite what you want to pretend.
re: Matt Math
"So you have to split it into two 16-socket hardware partitions..."
Sounds vaguely like a blade server to me. If you don't have the ability to scale, then why create the limited critter in the first place?
RE: re: Matt Math
Because for a lot of migration/consolidation projects there is value in being able to manage the whole even if it's two seperate hardware partitions. I'm not sure you'd want to talk blades, not from an IBM viewpoint (have to cripple the Power blades to keep within the poor power and coolling capabilites of their chassis) or Oracle's (what, after three or four complete failures, do they even do blades anymore?). And the scale is coming, which is why Allipoos is knocking it now whilst she still can.
HP does not have a 32 socket system....yes they will sell you two 16 socket systems and someday you will be able to link them together with wires to create a numa 32 socket system.
hmmm....they still dont have a 32 socket system.........clever confusion.
"obsurd"? Did you mean "absurd"?
"HP does not have a 32 socket system....." Aw, you really need to set the context better! I can quite easily point to the old Superdomes with last gen Itaniums, still available to buy, and you have a 64-socket option.
".....yes they will sell you two 16 socket systems...." No, they will sell you a 32-socket system that comes as two 16-socket hardware partitions, and will in future supply a kit to upgrade it into one 32-socket hardware partition. If that's confusing you then there's not much I can say, you just need to try a bit harder.
Interesting webcast next week
Looks like it might give some more insight into the Power roadmap