Intel unveils the Xeon E5-2600 series

Although the Sandy Bridge CPU microarchitecture has been available in laptops and desktops for over a year, it hadn’t yet made its way into the upper reaches of the computing ecosystem—until now. Intel has just introduced a range of Xeon processors intended for dual-socket servers and workstations, dubbed the E5-2600 series.

The silicon behind these new Xeons is a chip known as Sandy Bridge-EP. We’ve seen some of the potential of this chip in single-socket high-end desktop systems, where it has been shipping as a Core i7 for a while now. However, we haven’t seen the full extent of its capabilities in that application.

Sandy Bridge-EP builds on the same exact CPU core microarchitecture used in lesser Sandy Bridge chips, where its per-core prowess has been well established. This microarchitecture achieves substantially higher per-thread performance than Intel’s prior-generation Westmere processors, while adding support for new AVX instructions that double the width of vectors for floating-point math.

Content with the performance of the individual cores, Sandy Bridge-EPs architects have focused their efforts on integration of various sorts—of eight cores together, of smart power management policies, and of a staggering amount of on-chip I/O. This integration work promises huge benefits in performance, power efficiency, and system throughput compared to the last generation of Xeons based on Westmere-EP. The per-socket specs speak for themselves: 20MB of last-level cache, quad channels of DDR3 memory with transfer rates up to 1600 MT/s, dual QPI links, and 40 lanes of third-generation PCI Express connectivity—and all of those numbers double in a two-socket system.

Additionally, Sandy Bridge-EP has some impressive new features made possible by the increasing integration of various system components into the CPU silicon. For instance, the data-direct I/O feature allows data packets coming in from, say, a PCIe network interface to be deposited directly into the CPU’s L3 cache for processing, saving the overhead and power consumption associated with storing the data in main memory. Also, a revamped version of Intel’s Turbo Boost dynamic clocking scheme provides higher frequencies and longer residency at those higher clock rates.

All of this bandwidth and the I/O integration will, of course, require a break in socket compatibility. The E5-2600-series Xeons drop into a new LGA2011 socket that is part of a whole new platform, code-named Romley-EP. We expect every major server vendor—well, except perhaps SeaMicro—to introduce new products based on this platform very soon.

Here is a quick look at the various models of new Xeons, complete with Intel’s expected pricing.

As you can see, Intel has unveiled a pretty full lineup of products at various speeds and power envelopes.

The server flagship is the Xeon E5-2690, an eight-core part with a 2.9GHz base clock, a 3.8GHz Turbo peak with only one core occupied, and a 3.3GHz Turbo speed with all cores busy. The E5-2690’s 135W TDP is actually 5W higher than that of the corresponding Westmere-EP Xeons. Another model, the E5-2687W, pushes even further, into 150W TDP territory, which it can get away with because it’s a workstation-only product. That additional headroom allows the E5-2687W to run at 3.4GHz with all eight cores engaged.

We’d expect the E5-2660 to be one of the most popular new Xeons, since it’s the fastest model that squeezes into a 95W power envelope. With eight cores at between 2.2 and 3GHz, the 2660 should make for a nice power-performance compromise. Interestingly, so far, Intel has only two low-power variants of Sandy Bridge-EP on offer. The lowest-power option is the 2630L, with a 2GHz base clock, six cores, and a 60W TDP.

Damage Labs has been humming along rather noisily in the past couple of weeks as I’ve been preparing a full review of these new Xeons, including several of the models highlighted above. The initial performance results look quite striking, but there’s work left to be done before I can share a fully-baked version of them with you. In the interim, I’ll let slip one quick set of preliminary, experimental results. In our latest round of SPECjbb2005 tests with a new JVM, the Westmere-EP-based Xeon X5680 scores about 880K bops. The Xeon E5-2690, meanwhile, scores roughly 1.4 million bops. That’s quite a nice improvement from one generation to the next, I’d say. Stay tuned for a full review.

Comments closed
    • NeelyCam
    • 8 years ago

    BTW, Anand has a pretty extensive Ivy Bridge “preview” up now, with benchies.

    [url<]http://www.anandtech.com/show/5626/ivy-bridge-preview-core-i7-3770k/[/url<] NeelyCam's take: looks nice - I don't feel a need to upgrade from my 2600k. But, [i<]my god[/i<] look at that load power consumption!!

      • chuckula
      • 8 years ago

      When you look at the power consumption under load in the second pass of X.264 for Ivy Bridge compared to Bulldozer, and then note that IB is performing about 14.5% [b<]faster[/b<] doing it... AMD needs some serious help with Piledriver.

        • NeelyCam
        • 8 years ago

        What’s more is that the difference between idle and load is only about 50W! A 20-25W idle system is perfectly doable, which would mean at load it could run at some 70-80W.

        Or, rather, a dual-core system would be a perfect HTPC with passive cooling.

        Mmm… I’m tempted again. I promised I wouldn’t build another rig until Haswell is out, but a hobby is a hobby. Besides, I still have that spare 80GB 320 sitting on my shelf, just begging to be used..

      • Bensam123
      • 8 years ago

      Are we to assume from now on that every opinion will start with ‘Neely’s take’ and everything else you write as unbiased fact? >>

        • NeelyCam
        • 8 years ago

        Hmm.. should I do that? I like voicing my opinions, though, so writing “Neely’s take” on everything gets a bit tiring..

        How about if I abbreviate “NT:”..?

          • dpaus
          • 8 years ago

          How about: ‘speaking about myself in the third person is [b<][u<]O[/u<][/b<]stentatious and [b<][u<]S[/u<][/b<]nobby, but only half as much as some other people on here, so I'll just precede my comments with "OS/2:"....

      • OneArmedScissor
      • 8 years ago

      Load is nice, but I’m a bit disappointed with idle. All I really care for Ivy Bridge to accomplish is increasing laptop battery life in regular use, but it’s looking like this doesn’t have the impact that moving to HKMG did.

      I still think tri-gates are more of a necessity to keep shrinks working as expected than a significant technological leap.

      But of course, it’s not done yet, and even cutting 0.5w from laptops is tangible, so I eagerly await being proved wrong.

        • NeelyCam
        • 8 years ago

        I think the idle power figure is dominated by the test bench system.. My 2600k desktop idles at 22W. The IB system here idles even lower than the 2600k one, so it could help ny system break the 20W barrier.

        I have a feeling that modern CPUs do such a good job on powergating and power management that the power is mostly going somewhere else..

          • OneArmedScissor
          • 8 years ago

          I’m sure power gating makes the advantage less obvious. However, sizable parts of the chip, like the L3 cache and GPU, aren’t going to be gated, so there’s still [i<]something[/i<] to be had there. But it's all good, because I'm really more interested in the platform as a whole, and what OEMs make of it, rather than just the impact of the CPU. Will DDR3L be standard? Will they connect more things to the CPU's PCIe controller, which now has more lanes and might be a little more power efficient? These sorts of things are feasible, but are in their court. They've also been using 65nm for chipsets since Core 2, and I'm curious if they finally moved any of them to 45nm HKMG, but I have yet to see anything about that. The Windows power settings for laptops may also be more likely to trigger the new power gating features in deeper sleep states, like for the memory I/O, which Anandtech's test likely wouldn't reflect. These things could go a long way to cut the idle heat level in a laptop, which could save power two fold by allowing the fan(s) to run at an extremely low speed, or potentially shut off.

    • phileasfogg
    • 8 years ago

    Do any of the new Xeon E5 servers take advantage of the PCIe Gen3 speed available on the new Xeons? Perhaps 10GbE or Fiber-Channel controllers with native Gen3 interfaces aren’t ready yet? I know PLX had a press release today welcoming the launch of the ‘long-awaited’ (their term!) Romley platform, so I’m guessing that they have an axe to grind with this launch. PLX shows an application note with the 8732 Gen3 switch connected to two Fiber-Channel downstream controllers with x8-lanes each, at Gen2 speed. Upstream is x16-lanes at Gen3 speed connected to the Xeon E5 (?)

      • chuckula
      • 8 years ago

      1. All of them should have PCIe 3.0 because it’s built right into the CPU itself now*

      2. As for peripherals: That’s up to third party manufacturers to implement. Even PCIe 1.0 can be used to implement a 10 Gb ethernet MAC (assuming you use enough lanes). As long as peripherals are compliant, you should be in business.

      * Barring the outside chance that Intel is artificially limiting the lower-end models (grumble) although I haven’t seen anything in the specs indicating that this has occurred.

        • BobbinThreadbare
        • 8 years ago

        If I’ve done my calculations correctly, it take 5 lanes of PCIe 1.0, that doesn’t seem like it’s out of the question for a server board. Even accounting for overhead, an 8x slot should be plenty.

    • sschaem
    • 8 years ago

    $885 for a dual core CPU that turbo at 3.2GHZ and consume 80W! ? typo ?

      • chuckula
      • 8 years ago

      No.. but an extremely limited market niche that wants the ability to have a crapload of memory + very few cores. Supply & demand come into play since Intel will not be making a lot of those chips, the price is higher than you’d expect.

      Edit: I’m a little surprised that the 2-core part is not a successor to the Xeon 5698 (the dual-core 4.4Ghz clocked version of Westmere). I wouldn’t be surprised if in another 6 months Intel quietly releases an OEM-only dual core version with a 130 Watt TDP and a 4+ Ghz clock for niche clients.

    • bcronce
    • 8 years ago

    E5-2630L dual socket with 64GB of memory running FreeBSD9 with ZFS…. /drool

    next year maybe.

    • Bensam123
    • 8 years ago

    No one found Waldo? The 2637…

      • OneArmedScissor
      • 8 years ago

      The borked quad-cores have to go somewhere. They’ve left those to the server market since Nehalem. I wonder if it still has a quad-channel memory controller?

        • Bensam123
        • 8 years ago

        I have no idea what it even is, that’s why I was pointing it out. :l

        2 cores? That’s pretty much a desktop processor…

          • Krogoth
          • 8 years ago

          It is a low-level server chip, since the Xeon flavors of SB have ECC support. 😉

            • Bensam123
            • 8 years ago

            …but it’s more expensive the chips with better stats and lower power envelopes…

            • Krogoth
            • 8 years ago

            Xeon chips are tested with higher QA standards and come from cherry-picked yields that have lower power consumption.

            Besides, they are marketed towards enterprise crowd where $500+ price on CPUs is normal. 😉

            • Bensam123
            • 8 years ago

            I wasn’t comparing them to non-Xeon based chips, rather the chips in the above table. :l

      • Bensam123
      • 8 years ago

      Come’on guys, this was a good geeky joke!

    • bgoglin
    • 8 years ago

    “For instance, the data-direct I/O feature allows data packets coming in from, say, a PCIe network interface to be deposited directly into the CPU’s L3 cache for processing, saving the overhead and power consumption associated with storing the data in main memory. ”

    I don’t see anything new here. Intel Xeon have had DCA (Direct Cache Access) since about 2006 as part of Intel I/O AT. It’s already widely used by network interfaces and well supported by drivers.

      • Damage
      • 8 years ago

      Well, for one thing, PCIe wasn’t on the CPU, so the path to the cache was more convoluted. Also, I believe this feature doesn’t require I/O AT to be enabled and isn’t dependent on software support. I will ask for more clarification later, when working on the final review.

        • bgoglin
        • 8 years ago

        I/O AT is just the commercial word for several features including DCA and DMA engine, so yeah I/OAT isn’t required, but I/OAT is there anyway.
        DCA requires software support because you’re not going to prefetch incoming data in all L3 caches. You have to choose one of them, the one where the incoming data will be processed. Usually drivers update DCA hints when changing interrupt affinity.

    • chuckula
    • 8 years ago

    I think the TR review will be very revealing, but Anand did a decent writeup too. One line from the conclusion is really telling:

    [quote<]For those who are more price sensitive, the Xeon E5-2630 costs less than the Opteron 6276 and performs (very likely) better in every real world situation we could test.[/quote<] Ouch.... In the last go-around the Magny Cours systems were not necessarily the performance leaders, but at least they had a decent price-performance ratio that often beat the Intel equivalents by a good margin. Now we're seeing even the mid-tier Intel chips beating the high-end Bulldozers not only on performance, but also straight-up on price and on power consumption (which is also a related to the total cost of the system). AMD really needs Piledriver in the server ASAP.

      • NeelyCam
      • 8 years ago

      [quote<]AMD really needs Piledriver in the server ASAP.[/quote<] On that note, Anandtech had a good link to an ISSCC submission on the Piledriver resonant clock system: [url<]http://www.eecs.umich.edu/eecs/about/articles/2012/ISSCC_2012_Piledriver_final_submission.pdf[/url<] Interesting stuff.

        • ronch
        • 8 years ago

        Very interesting. Thanks for the link.

        • OneArmedScissor
        • 8 years ago

        This line is particularly noteworthy:

        [quote<]Figure 6 shows cclk and rclk waveforms from a full-chip clock simulation at 1.2V, 4.25GHz.[/quote<] If they're actually talking about the CPU core voltage, that implies AMD finally fixed the stock voltage issue they've had since the Phenom II. Everything from 2 to 4 GHz runs 1.4v, despite the fact that most of them work fine at 1.1v to 1.2v, just like Intel CPUs of the same speeds. The original Phenom did not do that, and I always found it odd that no one questioned them on this.

      • ronch
      • 8 years ago

      AMD has missed price/performance/watt by a mile. Given the die size of Bulldozer I don’t think they even think about dropping prices drastically to make up for the less-than-stellar performance and power numbers.

    • dextrous
    • 8 years ago

    “We expect every major server vendor—well, expect perhaps SeaMicro—”

    expect = except 🙂

      • Meadows
      • 8 years ago

      Shame on him. That’s a type of error usually reserved for the krogoth among us.

        • UberGerbil
        • 8 years ago

        [url<]http://www.youtube.com/watch?v=Wui-PNqJrxs[/url<]

        • Krogoth
        • 8 years ago

        [url<]http://www.youtube.com/watch?v=zPjUsu2-QMQ[/url<]

      • shank15217
      • 8 years ago

      you’re a fool, go lookup seamicro’s line up, they already have a SB based server platform which is already the most densest platform in the world… nice try

        • Palek
        • 8 years ago

        He was quoting Scott from the article above, so the target of your contempt is misplaced. What’s more, there is a link above to another news item in which we are informed that AMD is planning to acquire SeaMicro. You think SeaMicro will continue selling Intel based systems after the acquisition?

        You can remove your foot from your mouth now.

        P.S. Sheesh, this truly is the time of the ADD generation…

    • dextrous
    • 8 years ago

    I’m looking forward to this article!!! I wish you had more time to review more enterprise-target stuff.

    • [+Duracell-]
    • 8 years ago

    [quote<]the Westmere-EP-based Xeon X5680 scores about 880K bops. The Xeon X5-2690, meanwhile, scores roughly 1.4 million bops.[/quote<] Did you mean the E5-2690?

      • Damage
      • 8 years ago

      Yeah. Fixed. Thanks.

        • KinCT
        • 8 years ago

        One more to correct:

        “The E5-2960’s 135W TDP is actually 5W higher than that of the corresponding Westmere-EP Xeons.”

        E5-2690

    • chuckula
    • 8 years ago

    [quote<]Damage Labs has been humming along rather noisily in the past couple of weeks as I've been preparing a full review of these new Xeons, including several of the models highlighted above. [/quote<] We can't wait! 🙂

      • jdaven
      • 8 years ago

      “We can’t wait!”

      Speak for yourself only please.

        • chuckula
        • 8 years ago

        Oh it’s jdaven again.. are you going to deny the existence of the chips that Scott is testing here too just because ARM said that it might have a server chip available in 2014? Are you going to accuse TR of being in a giant conspiracy with Intel to sell non-existent silicon to the world?

          • tfp
          • 8 years ago

          Looks like some reviews are already up
          [url<]http://www.anandtech.com/show/5553/the-xeon-e52600-dual-sandybridge-for-servers/1[/url<]

      • khands
      • 8 years ago

      Agreed, exciting stuff in here!

Pin It on Pinterest

Share This