Slashdot Mirror
Intel Plans A Common Socket For Xeon, Itanium
stonedonkey writes "According to EE Times, Intel is planning a common system platform for the Xeon and Itanium by 2007, "creating a unified 64-bit motherboard with a new, one-size-fits-all socket." Intel's Jason Waxman says , "It has been something that customers have been asking us for for a while now...the reseller [currently] has to have an inventory of both boxes on hand." Feeling the heat from the competition, cutting losses, or just friendly customer service?"
The Inquirer reported on something just like this nearly a year ago. Of course this is actually confirmed with a date, but the Inq still has a bit more information.
The Itanium is crap isn't it?
I choose AMD and VIA mini/nano-ITX whenever possible, voting with $.
And my vote is to try and support VIA's innovation whenever possible, they are the only ones with the guts to break with the safe economics of the current form factors.
Glad to see that Intel is finally waking up and smelling the coffee.
Eg: how does a system bootstrap itself from power on to the point where the OS is loaded? That's the job of the code in the BIOS (or OpenBoot, or equivalent). Do you believe that x86-32 BIOS code is going to work to get an Itanium CPU loading the OS? The only way you'll get a generic BIOS to work is if there's an opcode that will allow a jump to a given address in one CPU, whilst just incrementing the instruction pointer on the other.
Look at Athlon. The Athlon bus was designed so that you could, in theory, plug an Alpha into an Athlon board. How many boards were made available to do this? Zip (that I know of, anyway.)
Intel are desperate to increase sales of Itanic (typo deliberate ;) -- they're hoping that by doing this, the economies of scale will make Itanic more appealing. Sorry, Intel; I'd say that this is the beginning of the end. Your fortunes were built on backwards compatibility, and it looks like that's now the millstone around your neck, dragging Itanium down.
First, the lead-in on Slashdot is silly. Intel has been planning this for a while. Yes, it is good customer service. They're not "cutting their losses" and this move makes sense even if AMD were to fold up tomorrow.
It's just simply too expensive to develop two different motherboard sets when you could leverage the increasingly similar characteristics of high-end Xeon motherboards and Itanium motherboards.
Also silly is the end of the article suggesting that Itanium will take over the world any time in the reasonably distant future. This is a strawman will no doubt ignite a frenzy of Itanium hating from various people (yes, we've all heard it before, Itanium is dead, Linus hates it, etc.). Besides, it doesn't reflect Intel's current clearly stated strategy, which indicates that we'll have both architectures for a very long time.
With Opterons coming in much less expensive than the Itaniums, moving to a common socket with the Xeons isn't going to help much.
On the other hand, with Opterons offering far better scalability and performance than Xeons, moving to a common socket with Itaniums isn't going to help out much.
Even on 2-way machines, Opterons show much better scalability than Xeons. As the number of CPUs increase, the Opteron architecture (when coupled with a supporting OS) allows it to shine more and more. With 8-way Opterons coming out fairly soon, Intel needs to come up with something fast - they're losing one of their most lucrative markets to AMD.
steve
Oh, you're not stuck, you're just unable to let go of the onion rings.
When will we see the return of common sockets on the desktop and server for both intel and amd processors? I'm starting to get annoyed with having to pay attention.
Intel would only do this if saving money was more important to them than giving people a credible reason to buy Itaniums instead of Xeons... and I do believe that's the situation at this point in time. Not a good sign for the future of Itanium.
If it were up to intel, we would all be using 120W processors that don't do shit... Intel did this for 1 of 2 reasons.
1. resellers were getting extremely pissed
2. they could save engineering, production, maintenance costs.
In general businesses try to "segment" their market, and when you vertically own the market, it is in your best interest to "make" everything, including different socket types.
Fuck Intel, they are feeling the heat because the P4 is turning out to be a piece of shit and the Opteron is kicking ass and taking names.
Granted Opteron's marketshare is nowhere near that of Intel's because all the OEMs don't wanna piss Intel off, but the big boys (IBM, HPUX, SUN) are selling Opteron systems because they know there is money and value in it.
Not that slaping together server hardware is that dificult (and I really don't know much about hardware design) but, wouldn't you want boards that are optimized for use with either Xeon OR Itanium? Wouldn't creating compatability on this hardware reduce its performance? Or is this a non-issue?
"Initial success, or total failure!"
remin8.com
And as a post pointed out next to mine, newer mainboards typically have plenty of other nice features.
If you're complaining about how much it costs to upgrade each time -- then don't. Especially these days, you really don't need the fastest and latest. I'm still using a 64MiB Radeon 8500 and I still don't see any compelling need to upgrade. (Although in my favor, I really don't care for Doom 3.)
Seriously - why would you want 128 bit addressing. Modern CPUs are already "128 bits" or more wide in many of the ways that count, such as SIMD registers and instructions. I'm not sure how many applications are working with data that would benefit from 128 bit wide integers and floats, but I'm going to make a guess at "not many."
The move to 64 bit addressing is being done mostly because we've run out of room on the 32 bit address space. I'm not dumb enough to say that we'll never run out of the 64 bit address space, but I think it's safe to say it'll take a fair darn while before we do. Remember that AMD's 64 bit CPUs only use 40 bit (or is it 48 bit?) physical addressing to provide support for terabytes of memory, so they've got quite a bit of headroom within the 64 bit address space.
This means you need to change your motherboard everytime you switch from one socket to another - big expense and something that makes u think a lot about which socket u want.
Well, I've had a PC for about 8 years now, and have upgraded a number of times. I don't recall ever keeping the same motherboard when buying a new processor.
True, I imagine that if you buy them regularly, then you may well "reuse" a motherboard once or twice. But personally, by the time I need/can afford to upgrade, a new motherboard is a requirement if I want to obtain a decent performance increase. My old one simply won't support the newer chips.
So, for people who do go to the trouble and expense of incremental upgrades, yes, I can see that not having to buy a new motherboard may be a factor. But I think probably the majority of people leave it long enough between upgrades that keeping the same one simply isn't an option. YM, of course, MV.
It's official. Most of you are morons.
Long ago, upgrading cpu with cpu upgrade kit from 386 to 486 and 486 to pentiums was quite a possible deal to boost an old machine up.
In 1997 I bought an Intel(!) mobo with Slot1/celeron buggy cpu. As was advertised these days, I hoped I will buy much better cpu some 2-3 years later instead of whole machine, because in past 15 years, as a developer I was buing a new machine for personal desktop every 2 years.
Unfortunately, no such upgrade was actually available, all "new" Intel processors were either completely different slot/socket or incompatible timig (mobo too slow). Again, in 3 years, I was forced to buy a new machine instead with some new stupid socket, which becomes obsolete today.
So, I do not believe Intel upgrade cpu/compatible cpu propaganda anymore. My new machine is AMD Opteron.
There you are, staring at me again.
The processors will assert different "core-type" lines, which will control which ROM is memory-mapped to the default EIP pointer at boot time.
Close. It would not make sense to have two or more ROM chips on the motherboard to support different architectures. So, what you do is take 1-3 lines from the processor or jumpers and connect them to the high order address lines on a larger capcity rom CHIP. 1 Line gives you two different architectures. 3 lines gives you 8 archs if you could ever get that many chip families in the same socket. Or take a chip about twice the capacity and use 4 select lines. Banks 0-8 of 16 would boot different processors and load microcode, chip specific code, and a java bytecode interpreter (not much in the BIOS needs to be very fast, except maybe the memory test). Then the whole top half of the ROM would be used for bytecode, with the 4 select lines being re-multiplexed as regular address lines once the firmware got to a certain point. Even better, a trivial change to the CPU cores could make them each start at different addresses in a jump table at the top of the ROM. Reserve the top 256 bytes as 16 different 16 byte locations. That should be more than enough for a jump instruction in any architecture, even with 64 bit opcodes with separate 64 bit addresses.
It really would be nice to have an industry standard processor bus which you could plug an intel, AMD, SPARC, or PowerPC chip into. It could be an interface with some optional pins. You could have a 128bit memory bus but only use half of it on some processors or motherboards. And you could have variable numbers of PCI-Express channels.
Failure to support different memory widths has been a traditional failing in PC motherboard designs. To keep it simple, consider the old 8 bit wide 30 pin SIMMS. Populating 5,6, or 7 of your 8 memory slots should have worked. In the case of 7, you would have had, say, 4MB of fast 32 bit memory, 2MB of medium speed 16 bit memory, and 1MB of slow 8 bit memory; you would use 4MB for code and the slower 3MB for cache or least recently used code pages (half way between ram and swap). I remember the 68020 family of processors was pretty good at handling this, there were lines the processor used to signal the requested width of the transfer and lines the external peripherals used to signal the actual width. These days with synchronous transfers the CPU would probably want to know the bus width of different memory banks in advance. Newer systems have similar issues with needing to install DIMMs in pairs to get full speed. It would be nice for people on limited budgets to be able to upgrade memory one bank at a time and the processor and motherboard at separate times. And do things like mix slow PC3200 RAM from your old motherboard and fast PC3200 RAM. Some minor OS improvements would be needed to allow the OS to recognize that not all memory is the same speed so slow memory shouldn't be used for speed critical pages. At some point, however, it is better to replace the whole CPU/memory/motherboard combination.
(sorry to add a serious comment to a joke)
The sockets diverged back in the days of Slot1. Intel didn't want to compete on the socket level, like back in the Socket7 days, so they tied things up in patents and such. That's what drove AMD to the SlotA, which they got from Alpha. Theoretically you could have motherboards into which you could put either a K7 or an Alpha, by changing the BIOS.
Actually, that same trick seems to be what Intel is trying to pull, here. But what seems and what is may be two different things. With today's market positioning, Xeon and Itanium are sold into different markets, and those markets would drive radically different types of boards.
So we have four possibilities:
1: Intel is trying to enable Itanium to creep downward into the Xeon market.
2: Intel is trying to enable Xeon to creep upward into the Itanium market.
3: Intel is confused, and trying to hedge their bets.
4: Beancounter have gained some ascendancy over the marketdroids, and have gotten fed up with the costs of 'socket differentiation.'
The living have better things to do than to continue hating the dead.
>The socket shuffle is no big deal, since by the time you want to upgrade your CPU you most likely will want a new motherboard too for new memory or other technologies. Motherboards aren't all that expensive regardless...
It is a big deal - one will be able to upgrade a Xeon system to Itanium with minimum effort.
Of course, the OS and the apps will have to be re-installed, but the box itself won't require rewiring and rack-mounting.
In data center environment the mobo itself doesn't contribute to performance a lot, but the key is in lowering the TCO, so it will matter.
And as the article said, having interchangeable mobos will make every Intel Xeon reseller also an Intel Itanium reseller at no extra cost.
The very fact that "mobos aren't all that expensive regardless" is the reason for this move. What is expensive is fscking with small volume of specialized parts, seinding technicians and engineers to Itanium training, etc.
The answer is simple: When you make a Pentium 4 chipset, you pay $4/unit to Intel. NVIDIA doesn't want to pay the licensing fee.
This sounds like Intel's Overdrive functionality that came on some 486 mbs...it had that extra socket that I *think* would take an early Pentium chip (or was it something special?) and would get you an effective speed boost without replacing your machine. I remember seeing a fair number of these mbs, but I don't remember if Intel even shipped anything to put in it.
I wonder if it backfired on them at all, I know a friend of mine had a mb with it and by the time Pentiums were all the rage he still had this 486 mb with this extra space that was nothing but promises and hype. I think he ended up getting an AMD-based machine because he was so angry at Intel.
x86 compatability is becoming increasingly irrelevant. Anything new for Windows is likely to run on .NET, and hence be CPU-agnostic, as long as you have a working .NET runtime on your platform.
If you truly believe that, I've got a bridge to sell you in NYC.
OLE, COM, Java, "web services", and half a dozen other "magic bullet" technologies have been repeatedly sold over the years as the ultimate in portability.
I don't think you will *ever* see big-iron or even medium-iron running any type of Microsoft OS unless it's an x86 chipset. Microsoft has neither the skills, the mindset, or a clue about big iron shops.
You are spot on with your analysis. I think the answer must be option 1 though. 2 can be ruled out because the Itanium has an insignificant market. 3 is a tautology for all enormous tech companies. And 4 is impossible :)
The reason Intel thinks they can use Xeon as a trojan horse for Itanic is Nocona, Intel's X86-64 clone of AMD's Opteron processor. Nocona will satisfy a lot of customers that are impressed by the promise of 64-bit computing but are too chickenshit to go with an AMD product. Once these folks already have a server with the proper socket an upgrade to the new processor will be a little easier.