Announcement

Collapse
No announcement yet.

CPUFreq Governor Tuning For Better AMD Ryzen Linux Performance

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

  • CPUFreq Governor Tuning For Better AMD Ryzen Linux Performance

    Phoronix: CPUFreq Governor Tuning For Better AMD Ryzen Linux Performance

    Our latest Ryzen Linux benchmarks are looking at the impact of the CPUFreq scaling driver's governors have on the performance of the Ryzen 7 1800X, including a look at the power consumption and performance-per-Watt when changing the governors.

    http://www.phoronix.com/vr.php?view=24288

  • #2
    Typos:

    Originally posted by phoronix View Post
    Ondemand - Commomly the default
    Originally posted by phoronix View Post
    Conserative - This is similar

    Comment


    • #3
      Originally posted by tildearrow View Post
      Typos:


      Fixed, thanks.
      Michael Larabel
      http://www.michaellarabel.com/

      Comment


      • #4
        Performance is pretty much what you'd expect. I've been more interested in difference in performance between different memory frequencies, as higher frequency RAM = higher frequency NB + higher frequency RAM = faster cache = faster cores.

        Comment


        • #5
          Originally posted by mmstick View Post
          Performance is pretty much what you'd expect. I've been more interested in difference in performance between different memory frequencies, as higher frequency RAM = higher frequency NB + higher frequency RAM = faster cache = faster cores.
          Obviously RAM speed tests with Ryzen will come once corrected firmwares are out so the memory can actually run at high frequencies and boot fine on Ryzen boards....
          Michael Larabel
          http://www.michaellarabel.com/

          Comment


          • #6
            The differences between schedutil an ondemand are razor thin, but the new schedutil governor does seem to have the occasional performance benefit and maybe even an occasional power savings. Good to see progress, even if minute and almost immeasurable. Might do some comparisons with ondemand in personal workloads.

            Comment


            • #7
              Has anybody done performance/watt comparisons of Ryzen with Broadwell-E? I haven't seen any so far.

              Comment


              • #8
                Originally posted by mmstick View Post
                Performance is pretty much what you'd expect. I've been more interested in difference in performance between different memory frequencies, as higher frequency RAM = higher frequency NB + higher frequency RAM = faster cache = faster cores.
                Also latency tends to go up with higher frequency, which will make it more interesting to investigate.

                Comment


                • #9
                  Is it possible, without recompiling the kernel, to completely override the governor with BIOS settings? And to do so in a way that the governor doesn't even run or cause interrupts?

                  Comment


                  • #10
                    Originally posted by mastercoms View Post

                    Also latency tends to go up with higher frequency, which will make it more interesting to investigate.
                    No, latencies (CAS, RAS, ...) are basically fixed and specified by the manufacturer in terms of nano/picoseconds in the datasheets. The times are then calculated in terms of clock cycles and programmed into the RAM (DDR1, DDR2, ...) controller.

                    As the programmed latencies can only be multiples of full (sometimes half) clock cycles, there is some rounding involved (you always have to round up). This typically leads to latencies (in picoseconds) going up and down when changing controller frequencies. The latency may be slightly higher at the highest supported clock rate compared to the second fastest, but it may also go down.

                    Comment

                    Working...
                    X