Announcement

Collapse
No announcement yet.

See How Your Linux System Compares To A $300 Broadwell-EP CPU That Lacks Turbo Boost

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

  • See How Your Linux System Compares To A $300 Broadwell-EP CPU That Lacks Turbo Boost

    Phoronix: See How Your Linux System Compares To A $300 Broadwell-EP CPU That Lacks Turbo Boost

    As I wrote about a few days ago, I'm in the process of my first Broadwell-EP Linux build and for it I had purchased the Xeon E5-2609 v4, a CPU that costs just $300 USD and has eight physical cores while a combined TDP of just 85 Watts, but it lacks Turbo Boost and clocks up to just 1.7GHz. But how does it perform?..

    http://www.phoronix.com/scan.php?pag...9V4-Benchmarks

  • #2
    I mostly don't care these days. Speedwise even old Phenom X4 955 is quite adequate these days, if we ignore power consumption for a moment, and even there GPU is much more "burning" issue. At least on Linux, there is much more to gain from SW optimisation than HW upgrade, so I don't choose HW in order to jump over a particular obstacle, but to hit some broad computing power target that I deem needed and then tweak/rework the SW.

    Zen benchmarks will be of interest to me since it represents qualitative jump ( new DDR4 with lower power consumption, higher freq / which is good for APU/ and higher density /16GiB-ones available, much lower power consumption on average, unified socket, new SSE and crypto engine etc etc.)

    Comment


    • #3
      Originally posted by Brane215 View Post
      Zen benchmarks will be of interest to me since it represents qualitative jump ( new DDR4 with lower power consumption, higher freq / which is good for APU/ and higher density /16GiB-ones available, much lower power consumption on average, unified socket, new SSE and crypto engine etc etc.)
      Agree completely, for most folks out there, unless you require a new machine right.now.today, you're better off waiting for Zen to hit the market. By all accounts, Zen should be a truly fantastic product. And even if you don't go with Zen, once its on the market, it will force intel to bring their pricing back down to earth, so it's a win-win for consumers. Competition usually is, after all!

      Comment


      • #4
        Originally posted by torsionbar28 View Post
        And even if you don't go with Zen, once its on the market, it will force intel to bring their pricing back down to earth, so it's a win-win for consumers. Competition usually is, after all!
        I really doubt this will happen so soon if at all, but sure a man can dream.

        Comment


        • #5
          Originally posted by torsionbar28 View Post
          By all accounts, Zen should be a truly fantastic product.
          There's no real evidence of that in any independent tests. As for the design of Zen, the cores are clearly an attempt to copy 2011-era Sandy Bridge chips while the cache design is more 2008-era Nehalem since AMD is using the same basic cross-bar switched cache that Nehalem had. This means that an "8 core" Zen is really more like two four-core chips that happen to be sitting on a piece of silicon and that have to communicate through the northbridge just like a 2 socket system would have to do.

          Obviously Zen will be a big step up from Bulldozer since Jim Keller was smart enough to throw out everything AMD had done with Bulldozer and started copying from Intel. It's almost impossible for it not to be a big step up.

          Comment


          • #6
            Originally posted by chuckula View Post
            "8 core" Zen is really more like two four-core chips that happen to be sitting on a piece of silicon and that have to communicate through the northbridge just like a 2 socket system would have to do.
            I thought that was what was done in most 8+ core designs too. NUMA?
            Also, can you link something about this?
            I only find stuff from totalbstech... I mean wccftech, which isn't exactly the most reliable source around. http://wccftech.com/amd-zen-cpu-perf...ouble-fx-8350/

            Comment


            • #7
              Originally posted by chuckula View Post

              There's no real evidence of that in any independent tests. As for the design of Zen, the cores are clearly an attempt to copy 2011-era Sandy Bridge chips while the cache design is more 2008-era Nehalem since AMD is using the same basic cross-bar switched cache that Nehalem had. This means that an "8 core" Zen is really more like two four-core chips that happen to be sitting on a piece of silicon and that have to communicate through the northbridge just like a 2 socket system would have to do.
              Maybe, but:

              - that "northbridge" is on the same silicon, which means that it can be as wide and as fast as needed.
              - this concept as it seems, have brought much better cache latencies. With even remotely smart job allocation, most of traffic will flow through cache of the "Zeppelin Quad" and traffic between quads will be much less. So it's very favourable tradeoff IMHO. Even if coherence demands bring extra clocks here and there, how much will this be felt at L3 latencies ?
              - If AMD stays true to tradition, with Zen we can expect all the goodies in base price, that otherwise carry premium sticker with Intel - like all virtualisation thingies, HT etc.






              Comment


              • #8
                the problem with the turbo-boost I had with my pentium g3258 anniversary edition (Hashwell) too. Here ist he script that write the correct values into the MSR registers to enable the turbo boost.

                #!/bin/sh

                # Toggle Turbo Boost for Ivy Bridge CPUs (should work for all newer Core)
                # Requires a fairly new Linux kernel (let's say 3.0+)
                # Written by Donjan Rodic, released for free use

                # check current real frequency with sudo turbostat -s -i1

                sudo modprobe msr

                # all_cores FOO
                # perform FOO(i) for each core i
                all_cores() {
                NPROCS=`cat /proc/cpuinfo | grep "core id" | wc -l`
                NPROCS=$(($NPROCS - 1))
                for i in `seq 0 1 $NPROCS`; do
                $1 $i
                done
                }


                # report Turbo Boost state on core $1
                read_tb() {
                ret=`sudo rdmsr -p"$1" 0x1a0 -f 38:38`
                [ $ret -eq 0 ] && echo "$1": on || echo "$1": off
                }

                # enable Turbo Boost on core $1
                enable_tb() {
                sudo wrmsr -p"$1" 0x1a0 0x850089
                }

                # disable Turbo Boost on core $1
                disable_tb() {
                sudo wrmsr -p"$1" 0x1a0 0x4000850089
                }


                if [ "$1" = "on" ]; then
                all_cores enable_tb
                all_cores read_tb
                elif [ "$1" = "off" ]; then
                all_cores disable_tb
                all_cores read_tb
                elif [ "$1" = "list" ]; then
                all_cores read_tb
                else
                echo "usage: turboboost.sh on | off | list"
                fi

                Comment


                • #9
                  Originally posted by starshipeleven View Post
                  I thought that was what was done in most 8+ core designs too. NUMA?
                  Also, can you link something about this?
                  [/url]
                  Yeah, never believe anything Wccftech says even if they literally copy-n-paste a real source and take credit for it.

                  Anandtech has a highly detailed two-part series about Zen's architecture that details everything I said:
                  http://www.anandtech.com/show/10578/...archy-revealed
                  http://www.anandtech.com/show/10591/...el-parallelism

                  As for the NUMA properties, Zen will clearly have strong NUMA characteristics that will not necessarily be easy to deal with. By contrast, an 8 core chip like Haswell-E includes a single shared L3 cache that is actually shared by all cores over an extremely fast ring bus interconnect. The only NUMA characteristic there is that some cache accesses can take just a little bit longer than others based on the ring stops, but those timing differences are basically rounding errors compared to true NUMA characteristics.

                  Incidentally, some people throw the word NUMA around like it's some wonderful property. It's not. It's a necessary evil in some large-scale systems and the OS + software have to be able to deal with NUMA in some applications like databases, but if you can avoid NUMA then you don't have to worry about all that stuff in the first place.

                  Comment


                  • #10
                    chuckula

                    Would that have anything to do with the fact that Intel has in-house silicon bakery while AMD has to go "classic way"- they just deliver files to GloFo,Samsung etc ?

                    This interconnect stuff sounds like a place where in-house switch and segment tweaking might get significant benefits over "LEGO" way ( syntesizing the logic on gate level).


                    Comment

                    Working...
                    X