As El Reg duly reported earlier today, Intel took the wraps off its long awaited and many times tweaked "Tukwila" quad-core Itanium 9300 processors for midrange and high-end servers. But let's take a look at the feeds and speeds of the chip itself and how the lineup compared to the prior Itanium 9100 series. There are five …
wait and see...
You felt the need to include the following in your analysis of Tukwila:
>(Wringing the full amount of performance out of any new chip does often require recompilation
>and other tweaking, however, and the Itanium 9300, with its radically different chipset and
>QuickPath Interconnect, will be no different in this regard).
This is equally true for the Power7, (going back to OOOE after the dalliance with IOE in P6) yet you saw fit not to mention it in your posts on P7... why?
And to the usual suspects who will post the usual "itanium FAIL" nonsense - let's wait and see a) what performance we see out of actual systems b) the price points of those systems. P7 may be fantastically fast, but for many people CPU performance is just not their issue any more, and the proprietary architectur of IBM system P is gonna look even more expensive over the next few years...
...in what sense is Itanium - or even x86 - non proprietary?
Errmm in the sense that you can buy x86 and Itanium systems from more than one vendor? Isn't that the defintion of proprietary? And in the x86 case, you can even buy a compatible chip from another vendor - AMD.
Ok, so by and large Itanium is proprietary to HP, but I think what people don't like about proprietary isn't so much the "vendor lock in", as the prices that usually come with it... lets see what sorts of prices tukwila systems hit the street at...
X5570 L3 cache
The X5570's 8MB cache is a L3 cache, it has 4 x 256MB L2 caches plus smaller L1 caches.
From the system uilders point of view the key thing about these new Itanium processors is that they are pin compatable with the new high end Xeons. So the likes of HP do not have to develop different hardware platforms for the two CPU lines.
"feeds and speeds"
....give it a rest with this one, it never made any sense to begin with.
whatever happened to the Itanic?
"pin compatable with the new high end Xeons" ??????
No they're not.
That was how they were orignally announced, and they both use Quickpath, but as of now they don't both fit in the same socket. Unless you can show otherwise. The next generation might, if we ever see them.
does it matter?
total pin compatability probably doesn't matter to most end users... how often are they really going to take out one processor and insert another??
what matters here is that the systems vendors can use volume chipsets in Itanium servers rather than custom chipsets - that makes overall systems pricing lower
if I understand the material released so far, HP will be able to produce up to an 8 socket/32 core Itanium server which could share almost every component with their x86 servers - that might not sound interesting technically, but it does mean that these systems benefit from the economies of scale HP have in x86 server manufacturing
what no-one seems to grasp here except for Intel is that the processor business these days is less about technology and more about volume/supply chain economics
>total pin compatability probably doesn't matter to most end users... how often are they really going
>to take out one processor and insert another??
Matters to the platform builders - and hence the price.
Always late and not worth the wait
The latest Itanium processor is again a disappointment.
Not sure why it is the 9300, maybe the DDR2 version which never made it to market was 9200.
Some glaring deficiencies are the 65nm fab and 1.73GHz. While Intel is bragging about 32nm in the beginning of the pitch they do not mention the nm on Tukwila.
While they brag about 8 threads per processor they are actually only saying it is still only two threads per core. For a 2 Billion transistor processor and 24MB of cache you would think it could have more than 4 cores.
The big deficiency is the 24MB cache. EPIC needs huge caches and Tukwila just cut the cache per core in half.
Intel gives numbers like 800% interconnect, 500% memory bandwidth, 700% memory capactity, but what matters is the chip and core performance. The chip performance is 2X Montvale, but the core performance is the same and only if you but the fastest processor.
I see SGI and Unisys have dropped out of the Itanium vendor list and the only people left are HP and people that are only looking to create a 4 socket mostly Intel box.
Strange seeing a big announcement for a chip that will not be in systems for 90days. Not sure the purpose except to say to HP's current customers that there is some hope.
Curious where the 2nd generation virtualization came from or was the 1st gen just on Xeon. Does any OS/software take advantage of any Itanium virtualization. All I have seen is partitioning and HP's IVM which is HP-UX and has 20%-50% overhead. True hardware virtualization is certainly needed.
"The 9300 is Intel's seventh generation of Itanium...There's one thing that people really demand here, and that's the ability to not have to do 'forklift upgrades' on a regular basis. They want investment protection." Kirk Skaugen, Intel vice president of the Intel architecture group and general manager of the data center group
==> Itanium 9300 is a forklift upgrade. There is no upgrade path from any prior chip and a mild socket only promise of the future.
clear example of a negative and make it look like a positive.
Curious how many QPI ports are on the chip and if it will suffer from the same chip hop problem that Nehalem EX will have beyond four socket. Reminds me of the AMD hyper-transport problem past 3 sockets.
RedHat RHEL6 - another ISV to drop support. Obviously RedHat is benefiting from the success of Power vs. Itanium.
Are their any virtualization offerings out there that can scale a virtual machine past 8 threads on Itanium? VMWare does not work on Itanium and HP's Integrity Virtual Machine last I checked had an 8 thread limit which makes it limited to one Tukwila chip.
The best they could do for a performance number is spec_int and fp but not actually put the numbers in the press release. I am betting Nehalem and Power are 2X to 4X Itanium.
Will there be a real benchmark any time soon like SAP or TPC.
Any vegas odds on Poulson 2012=>2014having the same schedule as Tukwila. 2006=>2010.
When does the Intel agreement with HP end?
Hey, I've seen this wall of text elsewhere! Tear down the wall!
SAP Power7 benchmarks
Well IBM have now published an SAP benchmark on Power 7.
85220 SAPs, thats a score of 2663 SAPs per core. Very impressive.
Power 6 had 1275 SAPs per core.
Nahalem has 2097 SAPs per core
AMD about 1000 SAPs per core
SPARC VII 685 SAPs per core
HP has not published any Itanium scores for a long time and the benchmark has moved on so can't compare scores.
relevance of per core score?
given that SAP don't price per core, what's the relevance of showing a "per core" SAPS rating?
SAPs per core
SAPs per core is usefull purely as a performance metric when comparing architectures. Also a number of SAP customers do purchase their databases from the vendor rather than SAP. SAP do price PI per core though!
Itanic yes but would be fun to play with
I actually would love to add an Itanic to my server farm at home just because esotic architectures and OS are fun to play with (stack machine cpu anyone?). Alas taking out a mortgage for the same cpu performance as one can get in a few commodity dell boxes makes this a nonstarter.
I bought one on eBay for $300. He must live in a tarpaper shack.
RE: Itanic yes but would be fun to play with
You could always go on eBay and buy a secondhand rx1600 or rx2600, either will run the current release of hp-ux 11i v3. Old generation yes, but great for a little playing with hp-ux (or RHEL, or SLES, or OpenVMS, or Windows). The problem is most of them have had their disks removed for secure wiping, so you need to find some hp disks and then get a copy of hp-ux (I never said it, but all new hp-ux servers leaving the factory seem to go out with a free copy of the Base OE in the box, even if they've been installed with a higher OE bundle.......) My other warning is they are rather noisy as they are designed for racks not desks, and the deskside kits for the rx26x0 range are hard to come by.
Blade doubts? Oh, come on! Try a little objectivity, maybe?
TPM says: ".....At 130 watts, it is going to be tough to put two of these Itanium 9310s on a single blade, and it would seem to be impossible to cram two of the Itanium 9340 (1.6 GHz) or 9350 (1.73 GHz) chips onto a blade, considering that each chip dissipates 185 watts...." Really? Why? Is there some magic figure of Wattage above which only a miracle will permit such a loading? Maybe IBM or Sun couldn't engineer to these levels but then hp have always been much better in blades design than either.
"....That said, the performance per watt on the Tukwilas should be a little better than with the dual-core Montvales. The top-end Itanium 9150N had 24 MB of L3 cache for its two cores and burned 104 watts...." Ah, so you really didn't do the maths in the previous statement. So a current BL870c blade that holds four dual-core Itaniums, so eight cores, would only need two quad-core 9300s to hit the same core count but provide better CPU performance and better throughput with DDR3 memory. And seeing as the Tukzilla quad-cores only draw 185 Watts for four cores whereas the old 9150N would need 208Watts for the same number of cores, ignoring the power chewed up just by the extra motherboard space needed for the two extra sockets, I'm guessing it will be quite easy for hp to make a new blade that will hold two 9300 Itaniums and outperform the old BL870c, for less power. At a guess, I'd expect hp to use the extra motherboard space freed up by removing two sockets to be used for more RAM slots, so the next gen BL870c (BL880c???) will also get an even bigger memory footprint.
Dear oh dear, TPM really didn't put much thought into that one, did he! Oh maybe he just did, but he was thinking more about the IBM story than the reality....
Itanium is dead
In my company we have migrated our Itanium servers to POWER7. As everyone else has done. This Itanium chip is the very last. Then, no more Itanium. Sad but true.
please return to planet earth!
so you've replaced itanium with a processor that was released 2 days ago- wow you work fast!
and given that yesterday Intel was talking about the next 2 generations of itanium 2 days ago as well, where do you get the "This Itanium chip is the very last. Then, no more Itanium" message from? I'm not here to defend Itanium, but _really_ some attempt at objectivity on the part of what is supposed to be a community of IT experts would be a reasonable start?
re: migrate from itanium to power7?
There's only one sane reason to do that: you've already verified that you absolutely cannot run 4 socket x64 Linux for your midrange apps or whatever you call your *nix. Not just one app. For that one mega app, go ahead, buy a big IBM.
But on a large scale, it is not sane. If you're undertaking application refreshes across the board, jump now to x64 Linux and save yourself a boatload of money over AIX. (I truly cry if anyone is running predominantly Linux on Power. Although it is better economics than Linux on z/OS, I suppose. Actually got a marketing email from IBM regarding that...)
Gartner sees it, most of the world sees it: x64 has been bumped many times and will continue to do so because of the competition between Intel and AMD. They're churning out some rather good processors on a yearly basis. Meanwhile, IBM will keep making these monsters that no one can afford -- hardware or software-- to claim some stupid tpcc or other irrelevant benchmark record. L/HPARS on big iron are insane. Neat, but many times more expensive than buying 2 or 4 way systems.
Most people will get on with life without the big iron. IBM is fighting the good fight, and I fully agree that Power has a much more viable future than Itanium. But don't kid yourself-- both ships are going down. IBM is just trying to make some money on the way out, preserve some lock-in another 5 or 8 years, and hope against hope that AMD dies and Intel goes back to sleep, because that's the only way Power is going to be a compelling performance-per-dollar solution.
PS, Matt appears to be either an HP employee or a dyed-in-the-wool HP customer. Allison appears to be a former HP customer. :)
Not sure why el reg embraces Itanium?!
Tukwila came awfully late, SPARC64 is in four core already at leas a year, shipping in systems. As well as Intel is x86 is in four core for some time already.
Eight core SPARC64 is ready, which is shipping to supecomputers today. Fujitsu is waiting for competitors to push this chip to UNIX servers.
And IBM announced POWER7 systems not only chip as Intel. This is very very big difference. You have to wait at least 6 months to get servers, containing Tukwila chips.
And if Oracle decides to decrease core multiplier for SPARC64 cores, Tukwila would be even more pointless.
Its obvious that Intel lacks of interest in Itanium development, because there wont be any significant volume and in microprocessor business volume is everything.
SPARC64 VIII is canceled
"Eight core SPARC64 is ready, which is shipping to supecomputers today. Fujitsu is waiting for competitors to push this chip to UNIX servers."
Not true from a number of aspects.
SPARC64 VIIIfx is only for the HPC market and is for the Japanese government project.
ready and shipping is questionable..but it is not for sale and nothing more than a picture of a chip as big as some ones palm.
SPARC64 VIII (Venus) was canceled about 18 months ago and there is no plan to bring it back into plan.
Thank you for your support of SPARC but let's not lead people astray.
... and when Sun announced multi core - multi thread
Seen this a lot from Sun, jonks ago - multi core, multi thread. Fact is, Itanium by another name is still late.
I am sure there is some contact somewhere which will run out soon which forces Intel to keep on rolling Itanium-type CPU's. Could expect a final merger of Itanium-type cpu's and the 'standard' Intel type CPU's soon.
Now if we can get Oracle to up the game? Naw - forget I said that.
Is SPARC64 Venus canceled? I must have missed that. Where are links so I can read more about it? Or is this just the ordinary FUD here, by IBMers?
I haven't heard it ..
was canceled. But then again I've never seen a roadmap where it was in any conventional business servers. aka the MX000 series.
So I wouldn't wait for it if I was you.
RE: I haven't heard it ..
I don't think Fujitsu have definately cancelled future SPARC64 outside HPC, but they certainly haven't jumped in with both feet. The problem for Fujitsu is they haven't got any message out there to reassure the existing SPARC base that they will definately produce a next gen Mx000 series. After all, a migration from current Sun-badged SPARC64 Mx000 and UltraSPANKed to a new generation of SPARC64, running the same binaries, would seem the lowest risk option and should be a winner with those companies stuck on Slowaris. I was expecting Larry to announce something positive when he did the Sunset victory dance the other week, I can't believe he's capitulating the high end so quickly. But then Larry never really wanted the Sun hardware business anyway. It looks like Larry's plans revolve around clustering lower end gear such as Niagara and x64 to try and make something that fits the enterprise, though I suspect he's got more than fingers-crossed that hp will keep him in the high end in the meantime. You may have noticed that Larry has spent all his time badmouthing IBM, not hp.
IVM does not scale compared to PowerVM
IVM has a limit of 8 threads per virtual machine. Thanks for posting the link to confirm this:
===> "Support of larger VMs (max: 8 virtual processor core)"
each Integrity virtual machine can only scale to 8 threads or the equivalent of one Tukwila chip.
The number of VM's supported per core is not interesting when a Power7 4 socket box supports 320 virtual machines. What is important is any PowerVM virtual machine can scale to 128 threads based on weightings on that four socket box
PowerVM can scale any virtual machine to 64 cores today with 256 threads, soon to be 256 cores and 1024 threads.
The reason HP did not yet announce the systems is because they would have to admit they will only have one 2 socket rack based system. All systems above two sockets will be retired and customers will have to move to blades cobbled together to create scale up systems. The rx3600, rx6600, rx7640 and rx8640 do not have any rack replacements.
vpar and npar will not be supported on the Tukwila systems either.
HP's "blade everything strategy" is more like "blade only strategy".
This strategy does not make sense given the HP DL980 system (Nehalem EX) is an 8U 8-socket rack system with gluefull chips for scalability.
RE: "IVM does not scale compared to PowerVM" - LOL!
"....IVM has a limit of 8 threads per virtual machine. Thanks for posting the link to confirm this:
===> "Support of larger VMs (max: 8 virtual processor core)....." And how exactly is that an issue? Unlike IBM and AIX, hp has a carefully designed and integrated virtualisation and partitioning range, referred to as the Partitioning Continuum. The Continuum part is because it caters for needs from sinlge, sub-CPU instances, right up to SMP instances, across the Itanium range. Seeing as you obviously don't have a clue how IVM works, I'll try and explain and also how it plugs into the Continuum.
The hp approach is to look at what customers do with their servers - usually consolidation, or partitioning a system to allow more than one software stack to run without interfering with another. In the hp Integrity range, the smallest servers are single-, dual- and quad-socket designs that use single motherboards (blades or racked servers), and from there on up they use four-socket cellboards linked with switching backplanes (from the 8-socket rx7xx0 range right up to the 64-socket Superdome SD64). IVM is sub-CPU virtualisation.
Listen carefully now, the next bit is important. As the single motherboard systems currently scale to eight cores, having IVM also scale to eight virtual CPUs is a perfect match. Above eight cores, hp has a couple of options. The first is based around hardware partitioning with a cellboard being the unit. Essentially, you can plug one or more cellboards into one server instance in electrical isolation from the other cellboards in a server. In the Superdome, that means you can create up to sixteen four-socket instances, each running it's own and completely separate OS instance (and you can mix hp-ux, OpenVMS, Windows and Linux in the one frame in separate hardware partitions if required). Or you could lump the cellboards together in combinations to suit your project. For example, if you were consolidating four quad-socket rx6600 servers and four octo-socket rx7640s into one Superdome then they could be arranged as four single-cellboard hardware partitions and four dual-cellboard partitions. Just in case you ever do get around to reading any hp manuals, hp refer to the hard partitions as npars (don't ask me why!).
Just to make npars more granular, hp also offer the ability to turn Itanium cores on or off using TiCAP (which stands for Temporary increased Capacity or something like that). This allows you to turn on CPUs when required, called being "active" (such as when meeting the end-of-year report run) and then power them off - "deactivated" - when not needed (such as when competeing against Power). So, an npar can be any number of cores from one to eight (or one to sixteen if Tukzilla cellboards stay as four-socket designs).
So, I hear you mutter slowly, what if you need a partition with more than eight cores but not a nice round number divisible by four, and you don't have a variance in demand that would make TiCAP the best answer? The answer is hp's next step on the Partitioning Continuum, called Virtual Partitioning or vpars (slightly more obvious than npars, I admit). The unit for a vpar is a single CPU core - you can then scale up to as many active CPU cores as there are in your hardware partition. If you wanted, you could split a Superdome into one vpar of a single core and make the remainder another vpar (don't ask me why you would want to do this, but then you IBMers do come up with some laughable system designs).
So, to recap - if all you need is up to eight virtual CPUs, then go IVM, or a single npar and turn on and off CPUs using TiCAP. If you need a number larger than eight cores and your hardware structure suits splitting the server into systems along one or more cellboards then go npars. But if you want to split into odd numbers of cores or a server size not in quads, then go vpars. As a last note, you can run npars and then IVM inside an npar, or npars and then vpars inside an npar, but you can't mix vpars and IVM in the same npar. Of course, the fun doesn't stop there, as hp tools such as Global Workload Manager allow you to run a solution across many npars or even separate servers all at once. In short, the most flexible and comprehensive partitioning and virtualisation package out there.
So, the point you are FUDing as hard as possible in reality has zero practical impact. I'm sure all the hp salesgrunts would like to thank you for this opportunity to show the customers how little IBMers actually understand about IVM, Inetgrity or Integrity's partitioning capabilities.
RE: IVM does not scale compared to PowerVM - part deux (or duh)!
Sorry, I was laughing so hard at your IVM non-point I forgot to point out the hilarity contained in the rest of your post.
"....The number of VM's supported per core is not interesting when a Power7 4 socket box supports 320 virtual machines...." So why mention that, and not mention that Power7 cores can only support half as many VMs per core as current Itanium? Or could it be that you wouldn't want anyone to think the Itanium cores are beefier than the Power cores and do more per thread? After all, it was weiner cores that relegated the Niagara design to their webserving niche, and you wouldn't want to happen to Power7, would you?
"....PowerVM can scale any virtual machine to 64 cores today with 256 threads, soon to be 256 cores and 1024 threads....." Hmmmmm, help my maths here, will you? You state a Power7 core will scale to eight threads, and each CPU will have eight cores, but that would mean you run out of scale with current PowerVM before you run out of cores, meaning PowerVM can't virtualise a top-line Power7 server! No wonder you want to stick to discussing a quad-socket server, as PowerVM won't virtualise the whole of say a Power7 equivalent of a Superdome. I'm guessing the "256 cores and 1024 threads" bit mentioned appears on the roadmap with as little certainty as Power8, i.e., "future". Oooh, busted again!
"....The reason HP did not yet announce the systems is because they would have to admit they will only have one 2 socket rack based system....." Really? Not what's on my NDA roadmap! And seeing as that came from hp, whereas your ramblings seem to come from your nether regions, I'm inclined to call your statement a cartload of male bovine manure!
".....All systems above two sockets will be retired and customers will have to move to blades cobbled together to create scale up systems. The rx3600, rx6600, rx7640 and rx8640 do not have any rack replacements....." LOL! Even if it were true, you contradict yourself - the rx3600 is a two-socket design! Duh! Try a little research before your next flight of fantasy. Then you "forget" to mention that even if it were true, hp blades go into racked chassis..... I suggest you call in sick tomorrow and spend the day doing a bit of basic research for your next dribbling post.
"....vpar and npar will not be supported on the Tukwila systems either...." And where did you get that from? Do you have an hp statement on that? Guess what - NDA roadmap says "Juanita is a big, fat liar!" Please, try and come up with something at least slightly believable.
"....This strategy does not make sense given the HP DL980 system (Nehalem EX) is an 8U 8-socket rack system with gluefull chips for scalability....." And here you really out-do yourself! Guess where the "glue" technology comes from? Yes, the same cellboard and switched backplane technology as Itanium! Intel has taken the QPI and cellboard technology from Itanium and pushed it down into the Xeon range to help it defeat AMD's Magny Cours. The exact same technology, only scaling to at least sixteen "nodes" (cellboards) and eight "glue chips" (switched backplanes), will mean Tukzilla in hp Integrity will still beat Power in the high end SMP market. But even better, since you brought up the DL980, you should also consider that hp has a target of sharing as many components between the ProLiant and Integrity ranges as possible, which means they will share disks, IO cards, system board components, and chassis! Yes, by pointing out the DL980, you just showed everyone EXACTLY what the next gen 8-socket Integrity server will probably look like! So a big thanks for helping disprove your own point.
/SP&L - it's like the Sunshiners, only without the pathos!
Do you know how HP-UX and OpenVMS compare? I dont know to much about them, but at my work, they praise OpenVMS above everything in terms of stability and clustering. Actually, I have been curious on OpenVMS, because the operations personel, sysadmins, etc all praise OpenVMS so much. It's extreme reputation must mean something, right?
And, I can not run OpenVMS on x86, right? In no way? Not via emulator, or anything? Do you happen to know? I must use an alpha or itanium, right? And they are not cheap. Server hw never are...
RE: Matt Bryant
"Do you know how HP-UX and OpenVMS compare?..." There are a couple of Alpha dinosaurs here picking up their clubs at the thought of what I might answer! We do still have a number of OpenVMS servers still running with apps that refuse to die. In truth, I've never had to put hp-ux up against OpenVMS in a comparative shootout so I can dodge the question (and the clubs!). OpenVMS, like hp's own old batch processing solution - MPE - both had big followings in the UK and, I'm told, abroad. As an OS, OpenVMS seems to have quite a good future as users seem very loyal and hp have committed to further development.
"....And, I can not run OpenVMS on x86, right?...." Correct, to run the latest version of OpenVMS natively you need Itanium.
"....Not via emulator, or anything?...." Not used one myself but I'm told there are x86 emulators used for OpenVMS programming. According to Hoffman Labs - an OpenVMS consultancy - you can use emulators to run "OpenVMS on a Linux, Apple Mac OS X or Microsoft Windows box". Or, if you want the real deal, you can always buy one of the older generation Integrity servers such as the rx1600 or rx2600 on eBay and run OpenVMS on it.
Ok, thanks for your help. I will investigate in OpenVMS emulators and Hoffman Labs. And also look at the rx1600 and rx2600 on Ebay (hope it is not too big, I want a small computer). OpenVMS is much used in finance and other critical tasks. In my work I must learn a bit about OpenVMS, one of our largest systems run on OpenVMS. That system is world famous. I would really like if OpenVMS were released for x86, then I would dual boot. Anyway, thanks for your help and have a nice weekend! :o)
Or try a DS10 or DS20
If you don't need the very latest version of OpenVMS just yet but want to get a feel for the OS then you could get an old Compaq DS10 or DS20, which were desktop boxes and therefore a lot quieter. They use the old Alpha CPUs so they won't run the very latest version of OpenVMS. I hear they're also very reliable, but a bit pricey on eBay as they seem very popular.
One of the dinosaurs has pointed me to the following webby for emulators and other OpenVMS goodies, apparently there's a program to register for and then they post you an OpenVMS media set:
/I have to go buy the dinosaur a pint in return.
Great! Ive looked at those rx1600 and rx2600 at Ebay. I will look at those DS10 and DS20 too. I dont mind if it is an Alpha, Ive heard they are quite fast, much faster than an equally clocked x86. Not as fast as Itanium, but still, they are plenty ok.
And thanks for that link, so buy that dinosaur a pint from me while you are at it, from me too! :o)
An free OpenVMS emulator under Windows. I dont have to buy extra hardware.
I hope I can use the latest OpenVMS, but maybe it is only realeased on Itanium. I have to look a bit. Anyyway, thanks a lot! I wish you lots of beer and sex in the weekend! :o)