AMD goes dodeca-core in Q1 2010

Robert Hallock (Thrax)

August 26, 2009 11:19 AM ET in News, , , , , , , , , , , , , , , , , , , ,

AMD Server Roadmap April 2009Over the past few weeks, AMD has been talking up its enterprise transition to DDR3 in the form of the 12-core Magny-Cours CPU.

AMD has used the multi-part “Playing 20 Questions” series of blog entries to paint the chip’s details. The processor is a dual-die solution that shoehorns two hexa-core Istanbul cores into one package and packs it in with cache and a spider’s web of HyperTransport links.

While a 12-core x86 chip is in and of itself a marvel, it is just one piece of AMD’s big step to embrace DDR3 in the enterprise. It is a move that represents a complete overhaul of AMD’s offerings with new chipsets, sockets, and processors.

San Marino Platform

Taking up the ground floor, the San Marino platform is a 1-2 processor solution for entry-level servers or workstations. Because of the move to DDR3, San Marino adopts the new Socket C32, which includes extra pins for DDR3 memory controllers.

On the processor front, Lisbon will take center-stage throughout 2010 and into the start of 2011. It’s best to consider the chip as a DDR3 version of AMD’s existing Shanghai or Istanbul products with a little optimization sprinkled on top. As the slides indicate, Lisbon will be sold under the AMD 4000 series and will be succeeded by the 32nm Valencia when AMD gets its 32nm production in check.

Maranello Platform

While San Marino cleans up in the minors, Maranello will swing for the fences in the big leagues. Positioned for the 2-4p server segment, the transition to DDR3 once again makes the case for the new Socket G34, which is mutually incompatible with San Marino’s C32 socket.

On the processor front, Magny-Cours is being primed for a 1Q10 launch and will take the banner as big green’s flagship chip until 2011. In ‘11, AMD is set to turn the 32nm dial to 11 and will test it on Magny die shrinks in the form of Interlagos with 12 or 16 cores.

As a flagship processor, Magny is worth extra consideration, and we wanted to briefly summarize some of the chip’s facts:

  • Contains two six-core Istanbul dies.
  • Each die contains its own dual-channel DDR3-1333 memory controller.
  • Fabricated on a 45nm process node.
  • Employs Socket G34 for 2 or 4-way systems.

Perhaps most significantly, however, is that AMD has strongly indicated that the chip will maintain the 75W thermal envelope offered by Istanbul.

“Our goal is to hold the power consumption of Magny-Cours consistent with Istanbul and that will happen with lower initial core speed and some new power management features,” says AMD PR rep Teresa Osborne.

The bigger picture

While AMD is working to dramatically overhaul its server offerings, existing Socket F products like Shanghai and Istanbul don’t hit the chopping block until the middle of 2011. In the mean time, Socket F will get a makeover when it joins the new San Marino and Maranello platforms on the AMD SR5690 chipset which brings AMD enterprise under the banner of a unified core logic.

Secondarily, the astute may have noticed that AMD appears to have exited the 8-processor game with the upcoming platform. This is no mistake. Market research conducted by AMD indicates that software is in a poor condition to scale out to the 96 cores that could be offered by an 8-way Magny-Cours system.

“The G34 systems planned for 2010 are expected to offer scalability to 4 sockets each with 12 cores per socket; that represents a total of 48 cores for a 4P platform,” said AMD Director of Biz Dev. “The challenge in moving beyond that level for many mainstream business applications can be that the scalability of the software becomes a bigger challenge than the scalability of the hardware.”

As it stands, a system packed with four Magny-Cours chips will crunch 48 independent processing threads. This level of computing density will be unrivaled until Intel introduces the equally impressive 128-thread Nehalem EX in the same time frame.

While the dust for this brewing battle won’t settle for quite some time, it’s clear that AMD is taking an aggressive stance and making the very best of its technology to craft a whole roster of exciting offerings.

13 Comments:

  1. Exactly what does one say about the possibility of 12 cores in a single package?

    HAL 9000 here we come?

  2. Exactly what does one say about the possibility of 12 cores in a single package?

    I say...neat. I hope people start to learn to write code that takes advantage of more than one thread soon.

  3. Unfortunately parallel programming isn't hit all that hard in undergrad courses and programming languages don't frequently make the task as straightforward as it should be.

  4. Yeah I agree. I'm not really sure what the solution is, though. Parallelism is hard to code, and I really don't see a new language fixing that problem. For me, most of my simulations are independent and can easily be parallelized, so it'll be nice for that sort of thing.

    All I ever really learned about parallelism was fork(); and how hard it is to do file I/O when you have multiple threads writing to one file.

  5. Some languages make the task of parallelism easier than others (like Erlang), but difficulty from implementation to implementation can heavily depend on the problem at hand.

  6. Or leave the core assignment up to the OS like it should be. Programs under normal situations shouldn't need to know or care how many cores there are. The OS should understand how many resources are at it's disposal and assign tasks to those resources accordingly. This whole situation we are currently running into with multi-threaded pcu's being neutered because programs aren't taking advantage of them is ridiculous.

  7. Uhh, the OS can't do all the parallel work for you.

  8. The OS can switch threads to different cores, the OS can't magically make a single threaded program multi-threaded though. They actually spent a fair amount of time dealing with multi-threaded programming in my college... well.. in one course anyway. It's really not that hard to do (depending on what exactly your code is doing of course) but it can be kind of tedious making sure you don't have any race conditions in your code.

  9. Im a programmer and to make it very simple for everyone to understand :

    Basicaly, the more your application use and lauch explicitly declared thread, the more it will benifit from more core.

    A thread is like a "channel" in wich you can launch a request or a method or whatever and it will not freeze you're application while its processing and you can continue to do other things.

    If the application use no thread at all, it will be very difficult for windows to make that application use more than one core. (Well, in fact even if you don't explicitly use tread, windows will use some while running but it will be on rare occasion).

    It takes some serious and heavy application to be able to program it in a way that it could use 12 core most of the time.

    To use 12 core you need to be able to split a single long task in 12 lesser task to run at the same time. (Wich is , kinda very hard to do, and probably wont last long)

    The other more common case is doing more than one thing at a time in the application. Saving and Encrypting some data while you're processing a search and sorting the results with the use of thread could use every core for a short time.

    For most users, more than 4 core atm is overkill, anything more than that is clearly for server. The only way you could benifit from 12 cores at the moment in a single desktop is running like 8 application and doing something in each of them at the same time, even then, there'll be time when some of your cores are idle if you have 12 of them :P

  10. I'll tell you where this will come in handy.... busy webservers running Apache with MPM/Worker. BTW: I LOVE MPM/WORKER.

  11. This can come in handy in almost any busy server :P On a busy server you never have enough core i think

    Never worked on an apache server tho, all our servers are Microsoft all the way =( Win2003 server and Win2008 server heh

  12. These are server processor's. Twelve cores in a consumer desktop will be a few years beyond the server application I would imagine.

    I'm running a quad with 8GB of RAM now, and its impossible for me to imagine anyone needing more multitasking power in a Windows 7 environment, at least until the applications start taking better advantage of the reserve capability for enthusiast desktop's.

    Servers though, that's a whole other ballgame. Keep in mind that those that preach the value of the future cloud will tell you that server CPU power is going to become king, and less and less will people depend on local hardware for computing as the broadband pipes open up and the cloud becomes a reality. I think its further off than some, but one day the CPU war may be exclusively fought inside of server farms, and not on a per desktop basis. If that is the case, it looks like AMD wants to be well positioned for that.

  13. This is virtualization for days. Load up several XenServers or Hyper-V a RDS farm on a quad processor system, I don't think 48 cores will be enough!!! Well, I guess having enough ram to support a full load might actually be a problem, server ram isn't exactly cheap.

Troll-free since 2003 ®