Announcement

Collapse
No announcement yet.

Intel P-State vs. CPUFreq Frequency Scaling Performance On The Linux 5.0 Kernel

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

  • Intel P-State vs. CPUFreq Frequency Scaling Performance On The Linux 5.0 Kernel

    Phoronix: Intel P-State vs. CPUFreq Frequency Scaling Performance On The Linux 5.0 Kernel

    It's been a while since last running any P-State/CPUFreq frequency scaling driver and governor comparisons on Intel desktop systems, so given the recent release of Linux 5.0 I ran some tests for looking at the current state of affairs. Using an Intel Core i9 9900K I tested both the P-State and CPUFreq scaling drivers and their prominent governor options for seeing not only how the raw performance compares but also the system power consumption, CPU thermals, and performance-per-Watt.

    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

  • #2
    These graphs would be so much easier to read if they contained percentages, i.e. you choose a certain governor as baseline (100%) and show other results in comparison to it.

    Comment


    • #3
      Hmm, a bit disappointed that there isn't an idle graph as well (or something like very light browsing), since that would be the best use-case for powersave. I would be curious to see if finishing a task early thanks to a more aggressive governor would lead to more energy savings.

      Otherwise, nice benchmarks as always, thanks

      Comment


      • #4
        Looks like CPUFreq powersave is the only real power saver, everything else is just more of the same.

        Comment


        • #5
          If anyone is monitoring their CPU with i7z and thinks that their vcore voltages are too high those are actually VID readout CPU voltage request and not the true vcore readout from what is set inside the bios. i7z is actually reporting VID voltages which are actually per core voltage request but the vcore voltage in the bios is only going to supply what it is set to supply. Don't know if anyone on here has ever researched or wondered about that. I thought I might share that. My 7700 nonk idles at 31c with 3 cores hovering around 4.1GHZ and one close to or hitting 4.2GHZ with P-State. Pretty aggressive turbo at idle. Full load on all 8 threads on all cores is a bit over 4GHZ. Its a 3.6GHZ base processor with a max single core turbo of 4.2GHZ. A vcore of 1.165 volts seems to work pretty well with this chip.
          Last edited by creative; 12 March 2019, 01:57 PM.

          Comment


          • #6
            Another interesting option would be to use userspace governor and then measure performance and power usage for specific workload and vary the frequency from min to max. This is the best way to find the optimum for your case. I hope to do something like this in a few weeks time for one of my test cases I've been working on recently.

            Comment


            • #7
              So in other words. If you have work to do, let it burn the midnight oil properly and go to real sleep as soon as possible. Everything else is just delaying the pain.

              Comment


              • #8
                This is only true if your algorithm and all of your data fits nicely into cpu caches.

                Otherwise you're limited by io, either to memory or to your storage. In such cases it does not make sense for cpu to spin empty at high speed, wasting power. You need to find the sweet spot, balancing power and time to result. This is especially important if you run your work at scale since it can make a difference between hundreds of kilowatts and a megawatt. Trust me, that is something noticeable on your power bill.

                Comment


                • #9
                  Originally posted by M@yeulC View Post
                  Hmm, a bit disappointed that there isn't an idle graph as well (or something like very light browsing), since that would be the best use-case for powersave. I would be curious to see if finishing a task early thanks to a more aggressive governor would lead to more energy savings.

                  Otherwise, nice benchmarks as always, thanks
                  I would also have liked to see idle power draw.

                  Comment


                  • #10
                    Originally posted by pegasus View Post
                    This is only true if your algorithm and all of your data fits nicely into cpu caches.

                    Otherwise you're limited by io, either to memory or to your storage. In such cases it does not make sense for cpu to spin empty at high speed, wasting power. You need to find the sweet spot, balancing power and time to result. This is especially important if you run your work at scale since it can make a difference between hundreds of kilowatts and a megawatt. Trust me, that is something noticeable on your power bill.
                    What you're talking about means very little if you completely turn off your system when you are not using it. I feel uptime records are for servers, its merely a bragging right for desktop users, Why let a system run if you are not using it or its not serving a purpose for being on?
                    Last edited by creative; 13 March 2019, 08:30 AM.

                    Comment

                    Working...
                    X