Rumor: Get ready for 12-core Opterons

AMD executives revealed during the latest earnings conference call that Shanghai, the company’s upcoming 45nm quad-core processor, will start shipping in volume in the fourth quarter of this year. According to a news post by DailyTech, though, the execs failed to mention that AMD is also planning six- and 12-core 45nm CPUs based on the Shanghai design.

Quoting an anonymous AMD engineer, DailyTech says Shanghai’s six-core cousin will be a native, single-die part code-named Istanbul. The chip will likely be a server model aimed at Intel’s Dunnington, which is also a single-die product and is scheduled to ship in the second half of the year. AMD won’t stop there, though—DailyTech goes on to say two separate sources have told it of upcoming 12-core CPUs made up of two six-core Istanbul dies on one package.

If this rumor is true, AMD will mimic the approach Intel adopted for its current quad-core x86 designs, which are all based on dual-core dies. However, where the Intel dies use the front-side bus to communicate with each other, a pair of Istanbul dies would presumably be able to talk directly via HyperTransport links.

Surprisingly, DailyTech quotes motherboard manufacturers as saying Shanghai’s six- and 12-core cousins will all be compatible with existing Socket 1207 (a.k.a. Socket F) motherboards. “Socket 1207+” mobos will reportedly be needed to take advantage of the future chips’ faster HyperTransport 3.0 connectivity, though.

Comments closed
    • AMDisDEC
    • 11 years ago

    I played with a Dunnington for a few weeks.
    It’s a sweet chip, but will be even better with QPI.
    Maybe this is why Intel pushed it out a quarter.

    • UberGerbil
    • 11 years ago

    It will be interesting to see how many channels of memory they need to hang off these things to not be bandwidth-constrained on many server workloads. As JBI says, the options are limited while remaining socket-F compatible, but if this pans out I wouldn’t be surprised if another high-end socket is on the horizon.

    • tfp
    • 11 years ago

    OH NOOOZZZ it’s not a NATIVE dodec-core processor. AMD is the SUX

      • Meadows
      • 11 years ago

      You mean duodecal.

        • tfp
        • 11 years ago

        I forgot to add the standard “But can it fold” post…

    • BaronMatrix
    • 11 years ago

    The funny thing is that Intel may lose it’s OCing ability when the mem ctrlr is on the chip versus in a chipset. That’s the disadvantage that AMD has now. You have to tweak three power planes to OC (CPU, HT, L3).

    Anyway, the numbers on Barcelona are up on SPEC and are looking good. if Shanghai really does have 15% higher IPC, there is a place for K10 on the Top 10 TPC-C list.

    It is supposed to have HT3 for Inter Processor Communications which gives 1hop 32 way servers.

      • UberGerbil
      • 11 years ago

      32? Depends on the number of cc links per socket. And 32 is a stretch. At some point you have so much cache involved that you need directories or other glue logic.

    • Usacomp2k3
    • 11 years ago

    I hate worrying about motherboard compatibility. Just give it an extra pin and not try to play BC that will only work on a percentage of boards.

    • no51
    • 11 years ago

    I hope these things are quad channel memory.

      • just brew it!
      • 11 years ago

      Unless Socket F has support for quad-channel (I don’t think it does), then no these chips will not support quad-channel memory.

    • dragmor
    • 11 years ago

    A HT link between the cores and each die having 1 memory channel out to the memory should make for a fast and easy approach.

    But they should skip the 12 cores and just do an 8 core.

      • Grigory
      • 11 years ago

      12 cores would be something. 🙂

      • charged3800z24
      • 11 years ago

      Well Intel will have 12 virtual cores.. 6cores with hyperthreading right? So this is AMD’s approach.

      • 0g1
      • 11 years ago

      Yeah, I’d be interested in an eight core. I wouldn’t mind if each die’s memory controller had to access its own separate memory, but it’d be nice if they could share.

    • maroon1
    • 11 years ago

    balablabala, AMD just like to talk a lot but they don’t show us anything

      • charged3800z24
      • 11 years ago

      lately they have not benn talking hardly at all.. They are just bring the crap out. And if you read.. it is “Sorces ” who say they heard the 12core etc….

    • Mitchell_F
    • 11 years ago

    AMD need to get their act together by dramatically improving K10 or releasing K11 before they push the boundaries because Intel could just come across and smash them which AMD does not want to happen. They should also see if they can get some of the fast switching transistors IBM used in power6 so they can just sell 4ghz processors because higher clock speeds will help them more than all the cores in the world.

      • Kurotetsu
      • 11 years ago

      As far as the server world is concerned, I’m pretty sure they’re more interested in multi-core than higher performance, since more cores = more virtualization and better load handling. From what I’ve been reading, AMD seems to have a better grasp on scaling then Intel (both in CPU cores and Crossfire), so that’ll be a big help for pushing this out. As well, they’ve always made everything backwards compatible, and while it may hamstring the chip’s capabilities, sysadmins will much happier with just swapping out a chip than having to swap out the motherboard, chip, possibly memory, and doing an OS reinstall.

        • Master Kenobi
        • 11 years ago

        I hate to break it to you but we sysadmins never swap boards or chips. We will buy a new server and move something onto the old server that isn’t too demanding for it. It’s very rare to see anyone in IT actually upgrade a server with newer boards and processors. It’s generally better to buy a new server.

          • flip-mode
          • 11 years ago

          Yep, for several reason: time, warranty, support, unknowns, probably some others I’m not thinking of.

            • BiffStroganoffsky
            • 11 years ago

            …purchasing decision(s) made by glorified receptionists… 😮
            Time wise, it is easier for me to bring down a server and swap out a couple of processors and bring the system back up than having to deal with the whole system setup and migration/replication process.

            • StashTheVampede
            • 11 years ago

            I’m going to ask this question: how often have you taken down a production server to upgrade the physical processors?

            Don’t get me wrong: I’m positive it has been done. How often have you done it? I’ve worked in IT shops for ~dozen years now, worked with fortune 500 companies and NONE of them have ever done anything physical with the machine OTHER than upgrading the amount of ram or adding more disk. The one exception was for a massive Solaris box where we took it down to add more physical CPUs.

            Servers are an example of computing “time stamps”. Most of the time they are never opened or touched until its time for them to retire or be moved to another location.

            • CheetoPet
            • 11 years ago

            Good point. I’ve only seen RAM & that was after a couple years of uptime. Support issues are gonna way complicate any CPU upgrade.

            • UberGerbil
            • 11 years ago

            You’re all correct. Essentially nobody swaps processors in working servers.

            However, pin-compatibility is still a win, because it offers cost and time-to-market benefits *[

            • charged3800z24
            • 11 years ago

            And if you are not some huge place and have a dual socket and want add vitualization and better benift from it… You might do a chip swap.. It isn’t completely out of the question…More so in a workstation.. Why spend thoudsand more for a complete package all over again and the added downtime and software loads. Atlest it is an option if you choose too, Right?

            • TravelMug
            • 11 years ago

            It happens sometimes. In one of the previous companies I worked for it happend twice. We upgraded the CPUs and added RAM to two Sun boxes. First was the production server running the lines which needed upgrade because it was chocking under the demand occasionally. The other one was also a Sun box running the Office tools for the thin clients under WABI.

            • BiffStroganoffsky
            • 11 years ago

            Twice, actually. They were the 940 Opterons and I swapped out for the dual cores when the sockets changed. Granted, that was an exception to the ‘rule’ around here as those were custom built boxes I put together on a shoe-string budget as a proof-of-concept that became production because the Compaq died as the deadline loomed. Nowadays, IT is mostly about CYA and having that factory warranty to rely/blame on when things go wrong. Those systems outlasted the Seagate hard drive warranty and were just demoted to development because another unit is taking over the hardware side of the business and run a server farm for people to rent virtual machines on.

            One of the reasons I chose the Opteron system to build was the fact that I felt AMD was going to give me a better chance of upgrading a socket 940 than Intel was considering how they went from socket to slot to socket. Not everyone’s definition of ‘production’ system is the same as a Fortune 500 business nor are their annual ‘maintenance’ budgets.

            • flip-mode
            • 11 years ago

            Depending on the server role, you have a point.

          • Jypster
          • 11 years ago

          In the high end X-series it was very common to swap out processors. These are designed around the fact that you could swap the KEG ( Contains CPU sockets and memory subsystems ) with relative ease. Each progressive model in the range was designed around this as it was a major selling point for Sysadmins. Buy a server with 4 processors now. Drop in another KEG down the track and you have 8 without having to purchase a new machine. Some machines allowed upgrading form 2 to 16 CPUs. Why replace ? having to reconfigure your Rack or stuff around with the connections, install your software, BTU changes new remote management software, new training etc . Your processors are a bit slow ? Buy 2 KEGs and replace both.

          Sysadmins that recommend complete system changes show that they did not research the equipment in the first place or did not foresee the companies IT demand growth. Maybe you don’t see upgrades any more as the level of technical know-how has been steadily dropping for years. Alot of admins only have the job because they know someone or were in the right place at the right time. I saw alot of this providing lvl 3 tech support… it was damn scary

      • bogbox
      • 11 years ago

      The 45nm is more then a shrink it has a new architecture 10.1 with bigger cache 6mb(only L3) so if they increase the memory controller to 3GHZ is better then a Nehalem
      the current bunch of phenoms have 1.8 GHz northbrige (except the 9850 with is 2.0 GHZ) so is a big deal .

      • 0g1
      • 11 years ago

      LOL fast changing transistors? You mean switching? The reason Power6 runs at a higher frequency is cos of its architecture, not the process they use (although it helps, but not that much!).

      A 12 core Shanghai would only come out a little bit before the Intel 12-Core Nehalem anyway. Should be closer performance than what we have now between Intel and AMD.

    • AMDisDEC
    • 11 years ago

    I wouldn’t hold my breath if I were you.

      • shank15217
      • 11 years ago

      oh look, its viral marketing guru AMDisDEC

        • AMDisDEC
        • 11 years ago

        Yeah, my forecast accuracy sometimes scares me too.

      • enzia35
      • 11 years ago

      Oh, cuz a lot of enthusiasts are desperately waiting to buy one of these for their main rig.
      I know I am /sarcasm.

    • bogbox
    • 11 years ago

    This is good news ,let’s hope is faster then Intel six core :))
    AMD is back in the game ! (at least at server part )

Pin It on Pinterest

Share This