Announcement

Collapse
No announcement yet.

AMD Ryzen Threadripper PRO 5965WX Memory Scaling Benchmarks On Linux

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

  • AMD Ryzen Threadripper PRO 5965WX Memory Scaling Benchmarks On Linux

    Phoronix: AMD Ryzen Threadripper PRO 5965WX Memory Scaling Benchmarks On Linux

    For those weighing whether to pursue the full eight memory channel configuration for the new Ryzen Threadripper 5000 series or starting out with just four or six memory modules, here are some reference benchmarks across four, six, and eight memory channels with the AMD Ryzen Threadripper PRO 5965WX 24-core processor while running Ubuntu Linux.

    https://www.phoronix.com/review/threadripper-5965wx-ram

  • #2
    Michael: since some workloads fared better with 6c, can you provide the memory clocks for the different configurations or are they the same?

    Comment


    • #3
      Originally posted by Anux View Post
      Michael: since some workloads fared better with 6c, can you provide the memory clocks for the different configurations or are they the same?
      Same XMP profile for all.
      Michael Larabel
      https://www.michaellarabel.com/

      Comment


      • #4
        I still cannot forgive AMD for making 5950X with just 2 memory channels

        Comment


        • #5
          This is actually pretty cool. So much improvement from quad-channel shows that the 5950X must be starving for bandwidth, with its simple 2-channel configuration. This makes me hopeful that DDR5 will unlock a lot of untapped performance, in the AM5 platform.

          For those who don't know, DDR5 has 2x 32-bit channels per 64-bit DIMM. So, a normal desktop machine will still have 128-bit wide data bus, but you'll actually get the parallelism advantages of a quad-channel setup.

          Comment


          • #6
            Can someone fill me in on why a result would regress going from 6 to 8 channels, but improve going from 4 to 6 channels?

            Why are some tests unable to scale? Does the test complete before all the channels are brought up for use?

            Does this reflect a power management activity or is it simply a memory switching activity?

            Comment


            • #7
              Originally posted by edwaleni View Post
              Can someone fill me in on why a result would regress going from 6 to 8 channels, but improve going from 4 to 6 channels?
              Perhaps it could relate to how the memory is partition among NUMA domains?

              Originally posted by edwaleni View Post
              Why are some tests unable to scale?
              Again, could be a NUMA thing or maybe the code in question just isn't sufficiently memory-bottlenecked.

              All of this is speculation, of course. Not meant as a substitute for actual knowledge.

              Originally posted by edwaleni View Post
              Does this reflect a power management activity or is it simply a memory switching activity?
              Interesting idea, but I doubt 2 extra DIMMs are going to significantly impact power budget, especially with a 24-core CPU in a platform designed for up to 64.

              Comment


              • #8
                Originally posted by V1tol View Post
                I still cannot forgive AMD for making 5950X with just 2 memory channels
                What is there to forgive? Then have 16 core chips with 2, 4 or 8 memory channels, take your pick.

                The 16 core is a bit starved in some workloads yes, but overall many people still opted for it and surely appreciate the option.

                Comment


                • #9
                  That's why I went the 2x7262 way instead of the TR. That and the ECC memory at the time.

                  Comment


                  • #10
                    Thank you for these benchmarks. They're useful.

                    Would be interesting to also run the same benchmarks on the 5950X and compare to really see the memory starvation, if present.

                    I kind of want to buy the 16 core 3rd gen TR as my primary workload would benefit from the memory bandwidth to core ratio. But I don't need to buy anything at the moment.

                    Comment

                    Working...
                    X