Intel unveils 10-core, 32-nm Xeon processors

Say hello to the latest additions to Intel’s server processor lineup: the Xeon E7-8800, E7-4800, and E7-2800 processors, all 32-nm offerings based on a new Nehalem-derived design that crams 10 Hyper-Threaded cores and 30MB of L3 cache on a single, king-sized piece of silicon. Intel offers variants suited for servers with two to 256 sockets, and it claims a performance lead of up to 40% over the Xeon 7500 series, its previous core-count champion.

Yep, that’s 10 cores. Count ’em.

The new silicon, codenamed Westmere-EX, has more than just preposterous core and cache counts to woo server makers. Intel says Westmere-EX also features new power-saving functionality, which "reduces the power draw of idle portions of the chip," as well as AES-NI and TXT instructions, which were not available in the Xeon 7500 series. AES-NI speeds up data encryption and decryption, while TXT (short for Trusted Execution Technology), "creates a secure platform at boot-up by protecting applications from malicious threats."

Intel’s Xeon E7 family includes 18 siblings, 10 of which have all of their cores enabled. Three of those tick along at 2.4GHz with a 130W thermal envelope, while the low-voltage E7-8867L hits a still-respectable 2.13GHz with a 105W TDP. The highest-clocked member of the family runs at 2.67GHz with a 130W envelope, but it has only eight cores and 24MB of cache. As you might expect, these new offerings aren’t cheap: Intel charges a whopping $4,616 for the top-of-the-line Xeon E7-8870.

Look for Xeon E7 servers soon from your favorite vendors, including Cisco, Cray, Dell, HP, IBM, Lenovo, NEC, Oracle, and many more.

Comments closed
    • alphadogg
    • 12 years ago

    50ft behind in a marathon isn’t “so far ahead”.

    • abw
    • 12 years ago

    What matters is performance , and in this respect , a 16C BD might crush this processor.
    Intel is just trying to take advantage of the few months before BD release to get as much monney as possible from this market.
    No doubt that the price will be severly reduced once BD show its tail…
    Btw : Interlagos will be in the 560mm2 range, so the intrinsical cost is
    about the same…..

    • OneArmedScissor
    • 12 years ago

    It costs way too much for that. Comparing core counts is dumbing things down too much and overlooking that there are so many platforms today because each has a specific application in mind, some more so than others.

    Again, the Xeon EX platform only exists at all because it’s filling a niche for a market that happens to have the money and need to justify it. They are specifically designed for use with extreme amounts of RAM, memory bandwidth, and CPU sockets.

    AMD opted out of this market completely and their current and forthcoming platforms just bridge the gap between several common use scenarios. They’d have “wiped out” Intel already with the quad socket X12s against the dual socket 6 core Xeons if that weren’t the case.

    • abw
    • 12 years ago

    It s the direct opponent of a 16C Interlagos since a 8C/16T will
    be undoubtly wiped off by AMD s top bin server CPU.
    Intel s slide can be viewed at Semi accurate forum.

    • OneArmedScissor
    • 12 years ago

    They probably made it because of requests from their customers that run ginormous databses, possibly even as a drop in upgrade for existing Nehalem EX users. Sandy Bridge EX is a little ways off.

    This is no Bulldozer competitor. It’s for companies with very niche requirements, who are willing and able to pay out the rear.

    • OneArmedScissor
    • 12 years ago

    The clock speeds are higher at the same power level.

    • abw
    • 12 years ago

    According to a confidential internal document that was running through
    the web, intel says that “depending on the perfs. of BD , we will add more cores”
    They say that they have the capability to release a 10 core cpu if needed..
    So we can safely assume that BD perfs is quite good.

    • NeelyCam
    • 12 years ago

    Could be. Or maybe this was the original plan all along… SB scheduled at Q3/Q4, and IvyBridge at Q1/Q2.

    So far it has seemed that Intel has been so far ahead in the race that they don’t really care what the guy 50ft behind is doing.

    • NeelyCam
    • 12 years ago

    ^ This. +3

    I should add that Intel went through the RDR at around the 45nm node (=the introduction of HKMG), which hurt their area efficiency. Others are expected to have to pay the same area penalties soon (TSMC @28nm, GloFo @22nm). This is one of those things that nobody in the industry can escape – it’s just the question of when do they tackle it.

    • Jigar
    • 12 years ago

    Good night… 😉

    • bthylafh
    • 12 years ago

    Can we expect a snarky review of this one?

    • indeego
    • 12 years ago

    60 ka-billion-gillion instructions per second on 7-zip, I’m guessing here.

    • dpaus
    • 12 years ago

    [i<]"winning!"[/i<] - So's Charlie Sheen. Just ask Him.

    • abw
    • 12 years ago

    Set aside a handfull of instructions , this processor is exactly the same
    as the preceding one , safe that it has been shrinked from 45 to 32nm.
    The only gain will be higher frequency, in that case 6% for the time,
    and the said 25% more cores which will not scale as well as frequency.
    Indeed, their own slide claim 25% better in INTEGER and 21% for FP.

    • Deanjo
    • 12 years ago

    It’s obvious that the right side was built by AMD of course.

    • Deanjo
    • 12 years ago

    Not at their claimed performance levels. A 40% increase in performance by only adding 20% more cores doesn’t jive. In a perfect world you would have a perfect linear scaling of 20% performance increase with increasing core count which is hard enough to do on it’s own.

    • dpaus
    • 12 years ago

    Thank you all, oh wise ones. I learn so much here…

    • shank15217
    • 12 years ago

    Intel is introducing this to take on bulldozer, which I think will be a bigger surprise than they anticipated. The timing is curious, it really wasn’t necessary as their sandy bridge based 8 core xeons will probably out perform these. Intel’s release of this chip means that they expect a core density war with AMD in two months.

    • demani
    • 12 years ago

    Huh? Are you saying its not possible for the cores to be more efficient than the previous ones?

    • shank15217
    • 12 years ago

    Its not two modules, its a single monolithic die, nothing asymmetrical about it.

    • UberGerbil
    • 12 years ago

    Yep, each generation is √(prev² / 2)
    So:
    sqrt(90^2 / 2) = ~65
    sqrt(65^2 / 2) = ~45
    sqrt(45^2 / 2) = ~32
    sqrt(32^2 / 2) = ~22

    There’s some rounding and other factors. One of the things we’re seeing with the transition to 22nm is that several factors (all related to the increasingly difficult physics) are driving the adoption of more restrictive design rules, which translates to less dense floorplans. So the density increase we can expect in the next transition will be less than we’ve seen recently, especially at first (things likely will improve as they gain experience with the node). If you want more details without trying to read something completely incomprehensible to non-specialists, Kanter at RTW has a [url=http://realworldtech.com/page.cfm?ArticleID=RWT031411013528<]nice summary of ISSCC 2011[/url<]

    • NeelyCam
    • 12 years ago

    You’re confusing “price” with “cost”

    • xeridea
    • 12 years ago

    You are off on your calculations. 45 to 32 is a 49.5% reduction. 32 to 22 is a 52.7% reduction.

    You square both numbers, so:
    45^2 / 32^2
    32^2 / 22^2

    Thats how each generation is ~1/2 the size, but the number is not half as small. its 32nm wide and 32 nm long… get it?

    • abw
    • 12 years ago

    25% more cores and up to 40% better perfs…Seems Intel is making fool of its customers…
    This is only true for the AES encryption since the relevant instructions were not present in the preceding 7500s…

    • sschaem
    • 12 years ago

    Thats almost 500$ per core… kind of pricey.
    Those core are much better then whats in Thuban (33$ per core), but I’m guessing electricity price and cooling bring this way above.

    But then, if you switch HW every year or two, you need to save 200 to 400$ per core in electricity a year to save any money.

    • bdwilcox
    • 12 years ago

    I seriously considered a variation of the meme (actually, I kind of did with my solitaire wisecrack) but ruled against it as my wisenheimer quota is running out for the day.

    • Steel
    • 12 years ago

    Why?

    • bdwilcox
    • 12 years ago

    You may be right. But I was under the impression that the rainbow effect came not so much from the microscopic grooves in the chips but from the topmost layer of the silicon wafer refracting the light. I read about that so long ago I don’t even remember where I read it; Byte magazine perhaps. It would be an interesting question for someone who intimately knows the process.

    • bcronce
    • 12 years ago

    Well, when they went from 45nm to 32nm (~29% smaller), they doubled the transistor count and still managed to make the cpu smaller.

    Going from 32nm to 22nm is a ~31% reduction in transistor size. So, again they should be able to double the transistors and manage a smaller chip.

    So, doubling a 6core cpu to 12cores while still being smaller/cheaper(cheaper once the process matures). The 6core goes for ~$1000 right now, so about 1/4th the cost of a $4k chip.

    So yes, 1/8th-1/4th.

    Nothing a person who can read and has a little background with computer couldn’t figure out. But who wants to wait around a year for a 12core when you can get a 10 core now!

    • Game_boy
    • 12 years ago

    Ivy Bridge EX will not be out until 2013. It will be the same price as today, so I think the 12 core variants will be half the price of today’s 10-cores at best. In 2013, not 2012.

    • Buzzard44
    • 12 years ago

    You should start an optimist club.

    • Krogoth
    • 12 years ago

    The only advantage that these chips have over the regular Gulftowns is having more cores/threads. 10-core Westmere-EX will mostly likely be thermally limited, if you wanted to OC them to any sort of degree.

    • bcronce
    • 12 years ago

    With 22nm coming out next year, I bet 12core Ivy Bridge CPUs for 1/8th-1/4th the price will be out.

    • UberGerbil
    • 12 years ago

    I’m sure a lot of modelling went into finding where the optimal core count was. Bigger dies are more expensive, and fewer fit onto a wafer, meaning defects have more effect on yield. Meanwhile they may have discovered a point of diminishing returns due to things like limited bandwidth from memory or on the internal bus or inter-processor snoop traffic or whatever.

    And power is a big deal, as Krogoth notes. You’ll notice these aren’t setting any GHz records.

    And then there’s planning for obsolescence, as OAS speculates, though I doubt that matters much to Intel: as long as they can keep pushing process technology, they’ll be able to obsolete their old models without explicitly holding back in a previous generation to achieve that.

    • indeego
    • 12 years ago

    Yawn, are you there? Wake me when this thing can run Crysis at $1/GHz.

    • Krogoth
    • 12 years ago

    TDP limitations.

    Gulftowns run pretty toasty with 6-cores when fully loaded. It is a surprise that Intel had manage to squeeze in 4 more cores with an more energy efficient variant. FYI, Gulftowns and these new guys are Westmere parts.

    • UberGerbil
    • 12 years ago

    dead memes are dead. winning!

    • OneArmedScissor
    • 12 years ago

    It’s not new, but it’s 1567, not the normal 1366 that standard Xeons use. 1567 is for eight socket boards, large amounts of RAM, and the CPUs have quad channel memory controllers and use the internal ring bus that Sandy Bridge has.

    • OneArmedScissor
    • 12 years ago

    “I wonder why they didn’t do it as a 12-core chip?”

    If they made it too compelling of a drop in upgrade for existing EX servers, that would tread on Sandy Bridge EX’s turf.

    • dpaus
    • 12 years ago

    Hey, how about a 5W mobile version in a 10″ netbook?? (I shouldn’t jest: this is probably about the amount of power needed to run a Star Trek Tri-Corder)

    • ew
    • 12 years ago

    Or use kvm.

    • Forge
    • 12 years ago

    If Intel sends me two, I’d be more than happy to lose the rest of my hair trying to find a proper motherboard.

    Heck, Intel, I’ll be your biggest cheerleader for a nice i7-990X.

    Who am I kidding, I’ll sell out for an i7-980X leftover!

    No? Nothing? I don’t love you either, Intel! I’ll go find a new best friend! 🙂

    • srilumpa
    • 12 years ago

    It’s actually to accelerate ed which is much more needed than accelerating vi (IMO).

    • bitcat70
    • 12 years ago

    I think it’s diffraction as in diffraction grating. Oil on water is actually an example of thin-film interference.

    • ronch
    • 12 years ago

    Hey, isn’t anyone gonna ask if this thing can run Crysis?

    • Steel
    • 12 years ago

    It depends on licensing. The free, Standard and Enterprise versions won’t use over 6 cores but if you spring for an Advanced or Enterprise Plus licence, you can use up to 12 cores per CPU.

    • derFunkenstein
    • 12 years ago

    That’s almost a full-size image of the die, too. :p

    • anotherengineer
    • 12 years ago

    Only $4600 w/o tax for an E7-8870, that’s expensive silicon, I guess that’s how Otellini makes 16 million/yr.

    But can it run the interwebs?

    • cheddarlump
    • 12 years ago

    Details on the chip are smaller than the wavelength of light used to photograph, leading to different interference patterns (combining of partial wave reflection/refraction) representing different colors.

    Same effect as looking at the bottom of a dvd/cd..

    • bdwilcox
    • 12 years ago

    Man, the power of these must be awesome! With 10 cores I bet I could run 10 games of solitaire AT THE SAME TIME!!!1!

    • GodsMadClown
    • 12 years ago

    For a moment I wondered if the TXT instructions were specifically for accelerating vi and/or vim.

    • bdwilcox
    • 12 years ago

    It’s actually light refraction from the surface of the chips, much like oil on water.

    • dpaus
    • 12 years ago

    Stangely asymmetrical; I wonder why they didn’t do it as a 12-core chip?

    Asymmetrical or not, very impressive engineering. Economically, not quite so much; a 4-socket system with 4- or even 6-core Opterons is still less.

    • blastdoor
    • 12 years ago

    20 core Mac Pro would be nice!

    • Ryu Connor
    • 12 years ago

    [url<]http://newsroom.intel.com/docs/DOC-1963[/url<] Larger image of the above die at the referenced link.

    • bdwilcox
    • 12 years ago

    Do these go in a new socket?

    • dpaus
    • 12 years ago

    Reflection of different wavelengths of the light used to take the picture, I think… Can anyone confirm?

    • TurtlePerson2
    • 12 years ago

    Where do the colors come from? Obviously they don’t use different metal layers in the left side as the right side, so why is one purple and the other green?

    • d0g_p00p
    • 12 years ago

    I would like to know this as well. I’ll be putting in 3 new ESXi boxes over the course of this year.

    • 5150
    • 12 years ago

    Double post.

    • 5150
    • 12 years ago

    Naturally I just ordered a new dual socket 6-core server for ESXi, but if I remember correctly, ESXi won’t run on procs using more than six-cores….????

Pin It on Pinterest

Share This

Share this post with your friends!