Announcement

Collapse
No announcement yet.

AMD Dual EPYC 7601 Benchmarks - 9-Way AMD EPYC / Intel Xeon Tests On Ubuntu 18.10 Server

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

  • #21
    There are some 2990 vs 7551 numbers available over at servethehome.com.

    Comment


    • #22
      Those are the old processors, aren't they? (Dual EPYC 7601)
      The 7nm are still coming out .... then it's really interesting.

      Comment


      • #23
        Originally posted by valici View Post
        Epyc domination... That thing should be illegal. Nobody should have that power.
        The right to bare cores.
        Last edited by msroadkill612; 16 October 2018, 09:05 PM.

        Comment


        • #24
          Originally posted by F.Ultra View Post

          Because I care not only about the single core performance.
          Who runs a 64 core server rack just for the single core performance?
          I have yet to see a usecase for such a machine.

          Comment


          • #25
            Originally posted by werpu View Post

            Who runs a 64 core server rack just for the single core performance?
            I have yet to see a usecase for such a machine.
            You could run the same program 64 times in parallel, with each instance working on a subset of the whole problem field.
            Maybe he also needs it to be reasonably fast or with guaranteed latency.

            Comment


            • #26
              Originally posted by mlau View Post

              You could run the same program 64 times in parallel, with each instance working on a subset of the whole problem field.
              Maybe he also needs it to be reasonably fast or with guaranteed latency.
              Then you dont run it for single core performance but for multicore performance :-)
              There always is a tradeoff between running a single core and running multiple cores at the same time, which is caused by power comsumption and thermal material limits and leak currents.
              So if you need this usecase simply look at the 7zip numbers or simething similar.

              Comment


              • #27
                So... You tested this CPUs in 22 disciplines, and AMD wins 21 of them?
                Can't believe my eyes.

                Can't understand TACC Frontera

                Comment


                • #28
                  Originally posted by werpu View Post
                  Then you dont run it for single core performance but for multicore performance :-)
                  Wrong.

                  Sometimes you want a single-threaded task to finish as fast as possible. Sometimes. CPUs are general purpose, you don't swap CPUs just for a one-off occasion, that's why they're general purpose. No amount of "doing other stuff at the same time" is going to help with that.

                  7-Zip is an interesting example. 7-Zip provides best compression ratio when single-threaded. So instead of using multiple threads for one archive, use single-thread for one archive, but encode multiple archives in parallel.

                  You'll still need single-threaded performance if you want one archive to finish pretty fast, but obviously you also want a lot of cores to do many at the same time.

                  Comment


                  • #29
                    Originally posted by werpu View Post
                    Who runs a 64 core server rack just for the single core performance?
                    I have yet to see a usecase for such a machine.
                    Except of course that I specifically wrote "not only" in the text that you quoted and answered to. So read the post you replied to again for your answer


                    Originally posted by werpu View Post
                    Then you dont run it for single core performance but for multicore performance :-)
                    And no one claimed that either. You run it for both the single- and multi-core performance. The current bencharks show the throughput of the processors, aka how many signs, reads, rendered scenes etc per second that the processor can process in total. What they do not show is how long each individual task takes to complete.

                    There exists use cases (and a lot of them actually) where you care both about the throughput and the latency so you have to weigh them both when choosing the hardware to run on but the current benchmarks only display one of these numbers which makes it hard to judge which one would be better for use cases where both aspects are important.

                    And both aspects are often the normal use case only that people don't know it and that the latency is often low enough for "normal" usage (but I'm in the sub microsecond latency business). E.g say that with one system you can handle 1M page visits per second on a web server but each page takes 20s to process and send to the end user (numbers inflated to highlight the problem) is this better or worse performance vs a different system that can only handle 100K page visits per second but where each page takes 500ms to process?!

                    For some use cases the 20s per page is actually acceptable and for others the 500ms might be the limit. If you then compare that with a third system that can only handle 50K pages per second but each page takes only 100ms then the second system might still be the better system for some use cases since 500ms is in the acceptable range and then handling 100k instead of 50k pages per second is more important than shaving off that extra 400ms.

                    So having access to both numbers are important to be able to determine exactly which system that will fit your specific use case.
                    Last edited by F.Ultra; 17 October 2018, 01:03 PM.

                    Comment


                    • #30
                      I've noticed the power efficiency myself testing the threadripper 2990WX against my dual-xeon. On the threadripper, the compute efficiency goes through the roof at slightly lower-than-stock PPT settings. AMD ships the chip a little on the hot side (~330W at the wall going full out). Reducing that to 250W and keeping the memory fabric at a modest 2666Mhz results in huge efficiency gains. It's good at stock settings, and even better when slightly detuned (in terms of compute power efficiency).

                      -Matt

                      Comment

                      Working...
                      X