Announcement

Collapse
No announcement yet.

IBM Opens Up POWER Architecture For Licensing

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

  • #31
    Originally posted by Ramiliez View Post
    Look

    But Power achieved those performance numbers with 96 threads against intels 64... those 5GHz chips dont look like 130W TDP which is AFAIK intels max TDP. So again you are comparing Apples to Oranges in single benchmark. Look corporations are most interested in performance in their workloads and performance per dollar and there IBM currently fails horribly with its 100% margins.
    well you are both not comparing aples to apples, Power main advantage over Intel is extreme scalability + infiniband, you won't buy Power system to run 4 sockets you will when you need to grow waaaaaayyyyyy beyond 4 sockets easily or when you need absolutely massive amounts of ram or when you wanna handle virtualization/isolation per process nodes be at socket or node level massively.

    With Power systems you just add boards with X sockets to a rack and plug infiniband you are done, do this as many times as you want practically without performance costs, Intel/AMD systems can't grow that big without custom desgined hardware and clustering techniques that have a bigger performance leak penalty.

    this are the scenarios where IBM Power Systems are actually cheaper than Intel/AMD in the long run[same apply to sparc], the point is they are designed for different scenarios.

    just ask the experts how much will cost you to put 2 thousand Xeon CPU to work together and then ask IBM how much it costs for the equivalent in Power and you will see what i mean.

    so the norm here is if you need to run many massive SAP[to stay topic] semi isolated system that each could handle million of transaction per hour and eat RAM like a turbo V8 an require centralized access to a massive terabytes long sets of Oracle/BD2 instances as fast as the technology can allow with rock solid stability and world class personalized support, Oracle Sparc/IBM systems are a no brainer

    if you have an SAP system single instance with a gigabytes long Oracle/Db2 database that process millions of transaction per day/week and would never map more than 512GB of ram, IBM/Sparc are waaaayyy to overkill[like killing a fly with a nuke] so here intel/AMD is a no brainer

    Comment


    • #32
      #jrch2k8

      I value your input

      To sum up your input i think IBM is actually trying to compete with SPARC in high-end not to compete with Intel

      As i understood it
      POWER advantages are good scalability ease of scalability its also good for supercomputer-like workloads due to x86 "leak penalty"

      Comment


      • #33
        Originally posted by Ramiliez View Post
        #jrch2k8

        I value your input

        To sum up your input i think IBM is actually trying to compete with SPARC in high-end not to compete with Intel

        As i understood it
        POWER advantages are good scalability ease of scalability its also good for supercomputer-like workloads due to x86 "leak penalty"
        1.) Yes intel has never attacked IBM Power market except the Itanium days, since its direct competitor is and was Sparc, at most intel can try in the lowest class of "Big Iron" system[my second example]

        2.) Yeap, Power/Sparc were desgined from ground 0 to be "Big Iron" CPUs while Intel designed 8086+ to be flexible CPU for PC and servers and its been forced to keep many silicon to provide compat with ancient stuff that time proved were very wrong designed[like memory mapping and the famous 640k/high mem/x87/etc].

        3.) x86 "leak penalty" is not exactly right, the point is "Big Iron" have many very very specific needs[segmentation for nodes/processes/containers, Accel encryption facilities, very extremely fast form of data bus, stacking infrastructure, FPU tier 1 precision, stability under extreme load handled by hardware, massive memory mapping thorugh busses instead of onboard, etc] that won't do any good to intel in other markets[maybe small supercomputers], beside this "Big Iron" systems are Pure 64 bits since decades ago[1993/UltraSparc III] and this architectures allow as many sockets/cpus as you want as long the internal inter bus have enough bandwith to handle them, this architectures are really efficient mapping memory and sharing the map among other processesor wherever the RAM is located.

        4.) supercomputer-like, is a bit wrong. "Big Iron" is a term used for heterogenous/simultaneous extreme loads while "SuperComputers" is more for homogenous/single extreme loads, aka "Big Iron" can handle thousands of different very heavy processes unrelated/related to each other in isolated/teamed form while "SuperComputers" try to solve few super extreme process as fast as possible, in this last scenario Intel/AMD are competitive since this super task can be very very parallelized and distributed among many many cheaper CPU/Servers that won't need to know what the other node is doing[most of the time] so this "leaks" are manageable

        Comment


        • #34
          #jrch2k8

          Thank you very much for extensive and detailed explanation.

          Comment


          • #35
            May I also add a #5?
            When building computers on the scale of thousands of cores, stuff like ECC through the entire pipe, not just the RAM will become essential, since you don't want one of your computations to go wrong and f*ck up the entire application. Especially when that server handles the pension system of an entire nation. This is just one of the RAS features that comes with IBM's high end server-plattforms.

            Speaking on pension systems.. point #6)
            IBM's high end processors, like the POWER6 and onward and the z/Architecture processors adds a decimal floating point units that handles currency computations without the rounding errors that regular binary FPUs are plagued with. When rendering 3D, film or games, accuracy is not that necessary, but calculating taxes, and wages and stuff like that, rounding errors can propagate and add up to very large sums in the end if this isn't done correctly and accurately.

            Comment


            • #36
              Originally posted by henriok View Post
              May I also add a #5?
              When building computers on the scale of thousands of cores, stuff like ECC through the entire pipe, not just the RAM will become essential, since you don't want one of your computations to go wrong and f*ck up the entire application. Especially when that server handles the pension system of an entire nation. This is just one of the RAS features that comes with IBM's high end server-plattforms.

              Speaking on pension systems.. point #6)
              IBM's high end processors, like the POWER6 and onward and the z/Architecture processors adds a decimal floating point units that handles currency computations without the rounding errors that regular binary FPUs are plagued with. When rendering 3D, film or games, accuracy is not that necessary, but calculating taxes, and wages and stuff like that, rounding errors can propagate and add up to very large sums in the end if this isn't done correctly and accurately.
              yeap i forgot those points but you are totally correct, in this systems ECC integrally integrated in every step is an absolute must[a system error propagated can cost a truckload of money in few seconds] and yes the FPU are slower that what you regurlarly find in your average CPU but is because (as well explained by henriok) FPU focus in mathematical precision instead of speed because with the sort of apps this equipment handle the lack of precision can become a very out of control nasty problem very fast for many vital industry sectors like financial[henriok put a very good example], advanced physics, engineering[you don't want a bridge unstable because rounding errors], goverment, military, applied sciences, space exploration, advanced signalling design, optics, among others.

              in this sense high end sparc is very competitive in the FPU precision but im not really sure how close to IBM tho

              Comment


              • #37
                Originally posted by uid313 View Post
                They should have done this 10-20 years ago.
                Now it is too late with Intel having almost total dominance, and ARM growing and into the server market now too.

                By the way, didn't they already do this long time ago?
                Wasn't the Power Architecture already opened and licensable a long time ago? http://www.ibm.com/developerworks/power/newto/#2
                Wasn't the firmware already open source long time? Open Firmware, SmartFirmware, SLOF are all open source already.

                What do they really mean?
                I thought this was already done long time ago...
                Yes they had a reference PPC motherboard design back in the mid 90's that could fit standard PC cases but they quietly had dropped it most likely due to the price which was out of the range of most enthusiasts.

                Comment


                • #38
                  Originally posted by henriok View Post
                  Speaking on pension systems.. point #6)
                  IBM's high end processors, like the POWER6 and onward and the z/Architecture processors adds a decimal floating point units that handles currency computations without the rounding errors that regular binary FPUs are plagued with. When rendering 3D, film or games, accuracy is not that necessary, but calculating taxes, and wages and stuff like that, rounding errors can propagate and add up to very large sums in the end if this isn't done correctly and accurately.
                  As long as you didn't have infinite Bits to store your (intermediate) Results, rounding hits. Are there Papers about this special FP Format available ?

                  Comment


                  • #39
                    I'm also wondering if this is true. Why would you add a special decimal format for money when you could simply calculate all money related stuff in cents with integers?

                    Comment


                    • #40
                      Originally posted by droste View Post
                      I'm also wondering if this is true. Why would you add a special decimal format for money when you could simply calculate all money related stuff in cents with integers?
                      well in economy there are a lot of calculations that require an FPU, is a bit more complex than count money using cents.

                      Comment


                      • #41
                        Multiple currencies exist in the world. We don't all use USD.

                        Comment


                        • #42
                          Originally posted by jrch2k8 View Post
                          well in economy there are a lot of calculations that require an FPU, is a bit more complex than count money using cents.
                          I'm not saying there's no need for an FPU (F as in floating)! I'm saying that you don't need a special fixed point unit.

                          /edit:
                          With "cents" I meant the smallest unit in your currency. I don't have USD in my wallet
                          Last edited by droste; 08-11-2013, 12:10 PM.

                          Comment


                          • #43
                            Originally posted by droste View Post
                            I'm not saying there's no need for an FPU (F as in floating)! I'm saying that you don't need a special fixed point unit.

                            /edit:
                            With "cents" I meant the smallest unit in your currency. I don't have USD in my wallet
                            well you have interest rates , money value proyections, devaluation, inflation, investment rates, taxes, deductibles, loans, etc. but at the end is a problem of quantity, this marginal errors won't be a problem for 100.000 transaction but for millions of transactions/per hour it grows to a big number, big enough to fear litigation that will cost you much more than just get an IBM power system to keep them under control

                            Comment


                            • #44
                              Originally posted by dibal View Post
                              As long as you didn't have infinite Bits to store your (intermediate) Results, rounding hits. Are there Papers about this special FP Format available ?
                              I think he means rounding errors when converting decimals into binary floating point numbers. More DFP info: http://en.wikipedia.org/wiki/Decimal_floating_point

                              Comment


                              • #45
                                Nowadays, looks like everyone open ups their IP for licensing .
                                Start with nVidia, then IBM. Who next?

                                Comment

                                Working...
                                X