Opteron 4100 series aims for the clouds

After telegraphing its intentions for quite some time, AMD followed through today with the official unveiling of the Opteron 4100 series of processors. The introduction of these new Opterons marks a bit of a departure from AMD’s usual strategy in server- and workstation-class processors, for several reasons.

The technology behind the latest Opterons isn’t substantially new. These processors were code-named “Lisbon” during their development, and the silicon is essentially the same as the “Magny-Cours” chips used in the Opteron 6000 series introduced three months ago. The big difference between the two is that the Opteron 6100 series employs two chips on a single package, cramming 12 cores and four memory channels into a single, outsized G34-style socket. The Opteron 4100 series drops into a smaller socket type, dubbed C32, and uses only a single chip and dual memory channels.

In fact, Lisbon isn’t terribly different from the Istanbul chips AMD introduced in the Opteron 2400 series over a year ago. Both chips have six cores and 6MB of L3 cache. In AMD’s silicon revisioning, the difference between the two is just a single stepping, from D0 to D1. The new stepping does bring a few changes, including support for an enhanced halt state, known as C1E, that the processor drops into when idle. Lisbon’s C1E state reduces power to the on-die memory controller and HyperTransport links via clock gating; AMD says the chip can wake up from this state in “under 50 microseconds.”

Notably, Lisbon silicon is different from the Thuban chips that power AMD’s Phenom II X6 desktop processors. Thuban chips have a further enhancement to GlobalFoundries’ 45-nm fabrication process—the addition of a low-k dielectric layer to reduce current leakage and thus cut power draw—and they support a dynamic clock speed feature known as Turbo Core that’s similar to the Turbo Boost capability in Intel’s newer Xeons. Lisbon and the Opteron 4000 series lack these features. In fact, Thuban is a later silicon stepping, known as E0.

Nevertheless, the introduction of the C32 socket enables support for DDR3 memory and higher HyperTransport speeds. The 4100-series supports both regular and the newer low-voltage versions of DDR3 memory at up to 1333MHz. Meanwhile, the two active HyperTransport links on the chip operate at up to 6.4GT/s, an increase over the 4.8GT/s transfer rate of Socket F-based Opterons. All told, bandwidth is up substantially for each socket. Peak memory bandwidth rises from 12.8 GB/s to 21.3 GB/s, and HyperTransport bandwidth jumps from 38.4 GB/s to 51.2 GB/s.

Although the socket type is new, the Opteron 4100 series will interface with AMD’s existing SR5600-series chipsets. Furthermore, AMD expects this platform to support CPUs based on the upcoming “Bulldozer” microarchitecture that is currently “planned for 2011.” Bulldozer promises big increases in integer math performance through the use of two tightly coupled cores that share floating-point execution resources. (This concept shows promise for the server space in particular, although AMD’s track record for delivering major architectural innovations in a timely fashion isn’t exactly spotless.)

If anything, AMD’s positioning of these new Opterons may be more novel than their technology. The Opteron 4100 series participates in a subdivision of the Opteron lineup in which the 12-core Opteron 6100 series serves most of the roles of past Opteron offerings. The 6100 series fits into two- and four-socket systems and collapses some of the distinctions between 2P and 4P servers, particularly in terms of processor pricing. The 6100 series also enables large memory configurations in 2P systems, thanks to its quad memory channels per socket. The 4100 series, on the other hand, attempts to move into some new market segments, as a look at the model breakdown and pricing will reveal.

 

By and large, these are cheap for server processors, and the lowest-end Opteron 4122 breaks some ground with its $99 price tag, which is under half the price of the most affordable quad-core Xeon on Intel’s price list. They also have relatively low power draw, with half of them below the mainstream 75W ACP band. The SE models with higher power and thermal ratings will remain exclusive to the Opteron 6100 series. Similarly, only the 4100 series has EE chips with ACP ratings as low as 32W.

These low-cost, low-power processors might seem to be suited for a certain sort of mission, and so they are. AMD will initially be aiming at a couple of markets with the Opteron 4100 series, the first of which is cloud computing or what the firm calls “hyper-scale data centers.” Such installations are a bit of a different animal than the traditional server market, because a single customer may be seeking a large number of systems tailored to its needs. In such cases, AMD may work with a custom design group at large server maker, such as Dell’s DCS group, to deliver a solution optimized for low power draw and high socket density.

In fact, AMD tells us it developed an Opteron 4100-based solution in conjunction with a “big web company” it could not name and a manufacturer in Taiwan. That system will serve as a reference platform for the Opteron 4100 series, and AMD expects the web company involved to be “very interested” in the finished products.

Above is an example of a power-optimized platform taken from an AMD presentation. This setup incorporates two complete dual-socket systems into what is presumably a rack-mountable 1U chassis. The total solution is claimed to draw just 125W, with a savings of 62W through various optimizations, such as a more efficient power supply and fans.

Separately, AMD cites an example of dual-socket server based on an Opteron 4164 EE that draws 57W at idle and 126W under load.

Naturally, one wonders how much business AMD can win by targeting cloud-type datacenter installations, but AMD is very quick to point out that we may see very little data about sales of this type, because custom design groups at major PC makers and their customers don’t tend to make much information public. Trouble is, the cloud computing push will be one of the Opteron 4100 line’s primary markets. The other market AMD expects to serve with these products is small and medium-sized businesses, which may purchase servers based on motherboards and barebones chassis from many of the usual suppliers, including Tyan, Asus, Gigabyte, MSI, and SuperMicro. The first of those motherboards and barebones systems should be available for purchase later this month.

What you won’t find, though, is traditional, high-volume servers based on the Opteron 4100 series from the likes of HP or Dell. Those simply aren’t coming any time soon, although AMD expects we may see some later in 2010.

The lack of design wins of this sort, coupled with the anticipated black hole of information regarding sales into cloud installations, raises some red flags about whether the market is really willing to accept AMD’s repositioning of the single-chip Opteron as a low-cost, low-power server solution. We do know that AMD expects to see a 70/30 split this year between the Opteron 6100 and 4100 series, respectively, with a shift to a 60/40 split in 2011. Assuming that happens, the Opteron 4100s may have managed to crack some new markets.

AMD seems to be taking a risk by eliminating the price premiums on its 4P products and repositioning its single-chip Opterons as low-cost, low-power parts, but the firm may have few alternatives given the current competitive picture, which seems to be rather difficult. Take, for instance, Dell’s recently published SPECjbb results, which show a dual 12-core Opteron 6176 SE system scoring 926K bops, while a dual six-core Westemere-EP Xeon system scores 928K bops. You’re looking at two smaller Intel chips outgunning four larger AMD ones with two additional memory channels attached.

Then again, raw performance isn’t everything, and AMD looks to be banking on that fact. Normally we’d resort to our own tests that consider performance and power draw in the context of price and positioning (and perhaps something else that starts with “P,” for good measure) in order to understand exactly what these new Opterons might have to offer. Unfortunately, we don’t yet have an Opteron 6100- or 4100-based system for testing, so we wait, semi-patient and ever vigilant, for our chance.

Comments closed
    • Mithrandir95
    • 9 years ago

    “Take, for instance, Dell’s recently published SPECjbb results, which show a dual 12-core Opteron 6176 SE system scoring 926K bops, while a dual six-core Westemere-EP Xeon system scores 928K bops. You’re looking at two smaller Intel chips outgunning four larger AMD ones with two additional memory channels attached.”

    Well to be fair you didn’t compare the JVM scores on that same link…

    154.6K for the Xeons, 231.5K for the Opterons. (50% faster). More in line with the Xeons 12 cores @ 3.33 GHz, compared with the Opteron’s 24 cores @ 2.3GHz.

      • accord1999
      • 9 years ago

      The difference in per JVM scores is caused by the Xeon system running 6 JVMs versus 4 for the Opteron.

      Giving the level of tuning that each submission gets, I’d expect 6 is optimal for the Westmere Xeons while 4 is optimal for the Magny Cours Opteron.

    • SomeOtherGeek
    • 9 years ago

    Just to set the facts straight: cloud computing = folding

      • Farting Bob
      • 9 years ago

      Cloud computing = anything at all…done anywhere other than the end user’s PC.

      • Krogoth
      • 9 years ago

      Nah, computing computing is terminal computing with a new name.

        • willyolio
        • 9 years ago

        cloud computing is terminal computing with /[

          • jinjuku
          • 9 years ago

          Wow, you mean the /[

      • Meadows
      • 9 years ago

      The other way around, my son.

        • poulpy
        • 9 years ago

        Not really no, Distributed Computing isn’t Cloud Computing.

          • Meadows
          • 9 years ago

          It is on the internet.

            • poulpy
            • 9 years ago

            Glad to hear that, still not the same thing.

            • Meadows
            • 9 years ago

            It must hurt to be wrong, do you feel it?

            • poulpy
            • 9 years ago

            I beg your pardon? What are you, 8?
            You’ve never been a joy to read but you seriously lost the plot here.

            Haven’t got all night on that so that’s the last of it for me but before you go to bed: §[<http://en.wikipedia.org/wiki/Cloud_computing#Comparisons<]§ q[

        • grantmeaname
        • 9 years ago

        you don’t really understand the concept of an equals sign, do you?

          • Meadows
          • 9 years ago

          When you read it out as “is” like I commonly do, then it’s not so very reversible anymore.

    • Anomymous Gerbil
    • 9 years ago

    How does “the anticipated black hole of information regarding sales into cloud installations” affect this chip, as opposed to any others?

    • WillBach
    • 9 years ago

    Did the math, looks like the C32 socket uses DDR3-1333. At the low clock rates I expect they’re not memory bound, which is a good thing.

    On a separate note, I wonder what optimizations let that 4-processor chassis run at only 125 Watts, when each processor has an ACP of 32 Watts? My guess is that if all four processors are running, a microcontroller pushes them into a lower clocks rate state.

      • Game_boy
      • 9 years ago

      ACP is like a worst case for typical workloads, much like AMD’s TDP is a worst case for any possible workload. So a given setup would use far less than ACP suggestsn if you just measure the power use/

        • NeelyCam
        • 9 years ago

        Westmere Xeon has an ACP of 10W

    • dpaus
    • 9 years ago

    Holy crap – a 16-core 2.2 GHz server running at 125W?!? and in a 1U case?? Unless I’m reading that wrong, that’s quite an achievement, especially if the total CPU cost of said box is just $396….

      • khands
      • 9 years ago

      I’m pretty sure all of these are dual socket only, so you’d max out at 12 cores.

        • jdaven
        • 9 years ago

        The picture shows 2 motherboards side by side with 2 sockets each. That’s 4 sockets unless I’m missing something.

          • dpaus
          • 9 years ago

          Yeah, dual motherboards, each dual socket = four sockets x 4 cores/socket = 16 cores.

          However, I realize that I /[

      • OneArmedScissor
      • 9 years ago

      It’s like putting two laptops together.

      This should have been done years ago. Desktops and servers waste absurd amounts of power.

        • eitje
        • 9 years ago

        But desktops (and to some degree, servers) have been cheaper to make until very recently (measuring cost of construction across 3 decades, let’s say).

          • OneArmedScissor
          • 9 years ago

          Yeah, because they don’t include formerly expensive LCD screens, batteries, and more involved cooling systems designed around almost no ventilation.

          They’re absolutely not cheaper once you’ve had them a few years eating 100w+ instead of 10w.

          There were numerous $300-400 65nm Athlon 64 and Core 2 laptops. Laptops used to be much more expensive than desktops, but it’s been years since the price void completely vanished, and all along, we’ve continued to be fed a bunch of baloney about new desktop parts being “more power efficient” when they can’t even hold a candle to a Pentium M laptop.

            • PrecambrianRabbit
            • 9 years ago

            Server and desktop parts aren’t doing so poorly in efficiency. They use more power than laptop parts, sure, but they perform better too. Granted the Pentium M was extremely efficient (that’s why the Core 2 was based on it), but there’s got to be a reason we don’t see Pentium Ms in servers… although we might see Atoms if SeaMicro gets its way ;-).

            • OneArmedScissor
            • 9 years ago

            “Efficiency” as it’s so often described doesn’t mean jack crap compared to wasted power over time. There is something wrong with desktops as they’re generally built, as they function almost identically to laptops and use several times as much power. I’m just going to leave it at that.

            • eitje
            • 9 years ago

            I said cheaper to make, not cheaper to operate. 🙂

            Also, you don’t need to sell me on the Pentium-M; I have a Shuttle SD11G5 that pulls less than 30W under load. It’s so awesome.

            • Anomymous Gerbil
            • 9 years ago

            Does anyone really care about power draw for home use?

            • bhassel
            • 9 years ago

            I care about heat & noise for home use…

            • reactorfuel
            • 9 years ago

            Power draw = heat = noise = /[

            • OneArmedScissor
            • 9 years ago

            You mean the only computer power use I actually have to pay for? Uh…is this a rhetorical question?

            • SomeOtherGeek
            • 9 years ago

            Yea, I care… My system is using about 150W as I’m typing and I think it is just too high.

            • Usacomp2k3
            • 9 years ago

            Not really.

            • flip-mode
            • 9 years ago

            I do.

            • eitje
            • 9 years ago

            low power, low noise computers are my hobby. 🙂

            • Saribro
            • 9 years ago

            My dorm room utilities bill went down notable when I started putting my system in standby when I wasn’t using it. I really felt the ease on my budget with some more thought through components/habits.
            So yes, I care.

            • OneArmedScissor
            • 9 years ago

            I know, I’m just ranting about how unnecessary additional cost of ownership is so often ignored.

    • khands
    • 9 years ago

    l[

      • Damage
      • 9 years ago

      Each Opteron 6100 processor is made up of two chips, each larger than an Intel Westmere-EP. So four total in a 2P box.

        • khands
        • 9 years ago

        Ooooh, my bad, nm then.

    • Krogoth
    • 9 years ago

    Interesting platform upgrade. To be honest, it is mostly for upcoming Bulldozer-based chips.

    Lisbon and Magny-Cours are just beta-testing the platform and stop-gap solutions.

    • ssidbroadcast
    • 9 years ago

    q[

      • NeelyCam
      • 9 years ago

      Have fun in Grammar Hell

      • Mr Bill
      • 9 years ago

      He could have been thinking… “from quite some time”, ‘ago’. Its not grammar nazi to help the article look better to the outside world; but rather, when you pick on a carelessly written post.

    • StuG
    • 9 years ago

    Good write-up.

      • WasF
      • 9 years ago

      +1

Pin It on Pinterest

Share This