Announcement

Collapse
No announcement yet.

AMD Talks Up Vega Frontier Edition, Epyc, Zen 2, ThreadRipper

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • AMD Talks Up Vega Frontier Edition, Epyc, Zen 2, ThreadRipper

    Phoronix: AMD Talks Up Vega Frontier Edition, Epyc, Zen 2, ThreadRipper

    Today was AMD's annual Financial Analyst Day where they revealed Zen CPU and Vega GPU details...

    http://www.phoronix.com/scan.php?pag...ncial-Day-2017

  • #2
    Extremely busy year for AMD.

    Looking forward to seeing how Naples performs in the data center. Naples should have a real performance per watt advantage over Xeon, and that is the ultimate metric in servers.

    Glad to finally hear *something* about mobile. Wondering if this means desktop APUs will also have Vega graphics. I would assume the answer is yes.

    Rumors are suggesting vega will be introduced with 3 SKUs ranging from $400 to $600. This price point is higher than I expected meaning Vega might be a better performer than I thought it would be. Either that or AMD is asking a crazy price for it. We'll find out very soon.

    Comment


    • #3
      AMD uploaded a brief promo video to youtube for Ryzen Threadripper: https://www.youtube.com/watch?v=VUuuuah2md8

      Comment


      • #4
        *sigh* Epyc? That's really the best they could do?

        "It'll be Epyc!" Not looking forward to all the puns. No thanks.

        Comment


        • #5
          Anyway, here's a newsflash. "Threadripper" is basically two Ryzen dies stuffed in a single package, connected via "Infinity Fabric", which is a custom protocol running over PCIe. Epyc is this, but with 4 dies in a package. And their inter-processor bus, in multi-CPU systems, is also Infinity Fabric.

          What does it mean? Well, it depends on how much you like NUMA. Because single-CPU systems will have the memory topology of 2-CPU and 4-CPU systems. Dual-CPU systems (though I don't know if you can do that with Threadripper) will then behave like 4-CPU and 8-CPU systems. So, thread/data locality will be crucial to getting good performance from it.

          The aggregate numbers are pretty eye-popping, but you'd do well to think of this as a density play. I'm looking forward to seeing some scalability numbers. I think Intel probably isn't too worried.

          Comment


          • #6
            I'm actually more interested in Vega. Although Nvidia's GV100 GPU will feature far more fp16 matrix multiply performance, its initial pricing actually leaves AMD a bit of room to shoot for price/performance parity. I expect early availability of AMD's solution will also be better.

            Comment


            • #7
              Originally posted by existensil View Post
              Glad to finally hear *something* about mobile. Wondering if this means desktop APUs will also have Vega graphics. I would assume the answer is yes.
              Yes (looks like you don't read enough Phoronix) - http://www.phoronix.com/scan.php?pag...-For-Gallium3D
              Michael Larabel
              http://www.michaellarabel.com/

              Comment


              • #8
                Originally posted by coder View Post
                *sigh* Epyc? That's really the best they could do?

                "It'll be Epyc!" Not looking forward to all the puns. No thanks.
                Epycspecially if everybody starts doing it.

                Comment


                • #9
                  Originally posted by coder View Post
                  Anyway, here's a newsflash. "Threadripper" is basically two Ryzen dies stuffed in a single package, connected via "Infinity Fabric", which is a custom protocol running over PCIe. Epyc is this, but with 4 dies in a package. And their inter-processor bus, in multi-CPU systems, is also Infinity Fabric.

                  What does it mean? Well, it depends on how much you like NUMA. Because single-CPU systems will have the memory topology of 2-CPU and 4-CPU systems. Dual-CPU systems (though I don't know if you can do that with Threadripper) will then behave like 4-CPU and 8-CPU systems. So, thread/data locality will be crucial to getting good performance from it.
                  It really depends on the task at hand, whether this is a problem or not. Example: assume we are operating with two or more files. You can easily split the tasks so that one 'real core' operates on one file. The tasks are totally independent. The same applies to all kinds of typical problems. Nowadays many CPU bound tasks (outside HPC) actually assume very high NUMA cost, much higher than it really is. It's up to the OS scheduler to prevent tasks from jumping from one NUMA node to another. The tasks that don't really fit in this model are often already fast enough. Even GPUs are very much NUMAs, e.g. imagine communicating with threads from another block in CUDA. People don't use 16-32 core systems with 32-64 threads to process parts of a single 2 megapixel photo. I'm not advocating NUMA here though. I'd really love to see something different, but that architecture doesn't exist yet.

                  Comment


                  • #10
                    Yeah, picking a name designed to appeal to teenage boys is a very odd choice for an enterprise CPU. But it's not like it could underperform bulldozer since it's at least a competent micro-architecture this time and you can just say "AMD Ryzen-based processor" or "32 core Naples CPU" and not one of those things like GIMP where you're just like "Hey, there's this graphics editor you should try, it's what I've been using and it's Free." "Cool, what's it called?" ".......uh .........Krita."

                    Comment

                    Working...
                    X