Announcement

Collapse
No announcement yet.

Trying The Configurable 45 Watt TDP With AMD's A10-7800 / A6-7400K

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

  • Trying The Configurable 45 Watt TDP With AMD's A10-7800 / A6-7400K

    Phoronix: Trying The Configurable 45 Watt TDP With AMD's A10-7800 / A6-7400K

    AMD Kaveri APUs feature a configurable TDP whereby users can opt to run their A-Series APUs with a lower power consumption and operating temperature but at the cost of slightly reduced performance.

    http://www.phoronix.com/vr.php?view=20764

  • #2
    Power measurement

    I think, it would be handy, if you could include some more real tests over time. For example "total power consumtion on: idling 5 minutes, compiling kernel, some graphics bechmark and again idling (different times to finish at same time)"

    Comment


    • #3
      While the option is nice, I feel like the need for this is pretty rare. If someone were to consistently lower the performance like this, they're better off just buying an A8 or maybe even an A6.

      Comment


      • #4
        CPU temperatures in benchmarks would be nice, but then the 45w configurable TDP would probably be worthless.

        Comment


        • #5
          Originally posted by GraysonPeddie View Post
          CPU temperatures in benchmarks would be nice, but then the 45w configurable TDP would probably be worthless.
          The CPU thermal monitoring doesn't work quite right for those APUs, as explained in other article.
          Michael Larabel
          http://www.michaellarabel.com/

          Comment


          • #6
            performance per watt numbers are wrong?

            I'm not sure these Performance/Watt values make sense in every case. While perf/Watt is better for 45W TPD (although you write: Xonotic was also more power efficient with the stock TDP.), it's worse for the benchmarks you measure the time to complete the task. I'm not even sure how you get "perf/Watt" for the "Linux kernel" and the "c-ray" benchmark.

            Naively, I'd do the following: To calculate something proportional to "performance", I'd take the inverse time ( N Operations take m seconds, therefore the performance is N/m). As I don't know the number of operations, I'll leave N=1. Thus, performance per watt is:

            Linux kernel:
            stock: S = 1 OP/(163.81s * 111.8W) ~ 5.46e-5 OP/s / W
            45W : T = 1 OP/(183.27s * 93.6W) ~ 5.83e-5 OP/s / W

            Or normalized to stock: T / S ~ 1.07
            C-Ray:
            stock: S = 1 OP/(58.64s * 109.1W) ~ 1.56e-4 OP/s / W
            45W : T = 1 OP/(66.73s * 91.5W) ~ 1.64e-4 OP/s / W

            Or normalized to stock: T / S ~ 1.05
            Thus, performance per watt is in any case worse for the stock tuning.
            Last edited by oleid; 08-13-2014, 01:30 PM.

            Comment


            • #7
              .... except the valley benchmark of curse! But that's not playable anyway

              Comment


              • #8
                Originally posted by oleid View Post
                Thus, performance per watt is in any case worse for the stock tuning.
                I fully agree with oleid; the task is to compile the Linux kernel:

                Stock TDP: 163.81 seconds * 105.5 W = 4.80 Wh
                45W TDP: 183.27 seconds * 87.9 W = 4.47 Wh

                So for one task (compiling the kernel), the 45W TDP setting uses less watts per hour, therefore it's more efficient!

                Comment


                • #9
                  Originally posted by oleid View Post
                  I'm not sure these Performance/Watt values make sense in every case.
                  They don't. People have been complaining that the formulas used for performance/watt are incorrect and return completely stupid and wrong values since the very first article of this kind was published some 2 years ago, but Michael just doesn't care, and prefers to have useless articles with wrong data and conclusions, since that probably makes him more money. At least people are commenting on the forums that the calculations are wrong, where if they weren't wrong, people would only have valid data to talk about. Who wants that, right?

                  There was even this case in one of the articles that was soooo obvious and yet he published the damn thing. It was like a cpu was consuming double the energy of another one while only providing some 10% increase in performance and the cpu with the higher energy consumption was deemed as the most efficient. Seriously... don't waste your time. It will never be fixed.
                  Last edited by devius; 08-13-2014, 02:36 PM.

                  Comment


                  • #10
                    Originally posted by devius View Post
                    There was even this case in one of the articles that was soooo obvious and yet he published the damn thing. It was like a cpu was consuming double the energy of another one while only providing some 10% increase in performance and the cpu with the higher energy consumption was deemed as the most efficient. Seriously... don't waste your time. It will never be fixed.
                    Yeah, i remember that and thinking "how the hell does michael think this makes any sense at all"?

                    Comment


                    • #11
                      Wrong, wrong, wrong

                      "Xonotic was also more power efficient with the stock TDP."
                      The chart above schows clearly the opposite!

                      Comment


                      • #12
                        Originally posted by drSeehas View Post
                        "Xonotic was also more power efficient with the stock TDP."
                        The chart above schows clearly the opposite!
                        Yep, and furthermore, the provided values show, that the effiency graphs of the last page are wrong.

                        Comment


                        • #13
                          When the test is more is better, the results of performance per watt are fine.
                          When the test is lower is better, the results are wrong.

                          Examples with 1 variable and 1 constant:

                          More is better test:

                          PC 1 : 100 frames per second / 50 watt = 2
                          PC 2 : 100 frames per second / 25 watt = 4
                          Correct.

                          PC 1 : 100 frames per second / 50 watt = 2
                          PC 2 : 50 frames per second / 50 watt = 1
                          Correct.


                          Less is better test:

                          Seconds/Watt(wrong)

                          PC 1 : 100 seconds / 50 watt = 2
                          PC 2 : 100 seconds / 25 watt = 4
                          Correct.

                          PC 1 : 100 seconds / 50 watt = 2
                          PC 2 : 50 seconds / 50 watt = 1
                          Incorrect.


                          Watt/Seconds(wrong)

                          PC 1 : 50 watt / 100 seconds = 0.5
                          PC 2 : 25 watt / 100 seconds = 0.25
                          Incorrect.

                          PC 1 : 50 watt / 100 seconds = 0.5
                          PC 2 : 50 watt / 50 seconds = 1
                          Correct.



                          The right formula is:

                          1 / ( Seconds * Watt) and normalize.

                          PC 1 : 1 / (50 watt * 100 seconds) = 0.0002
                          PC 2 : 1 / (25 watt * 100 seconds) = 0.0004
                          Correct.

                          PC 1 : 1 / (50 watt * 100 seconds) = 0.0002
                          PC 2 : 1 / (50 watt * 50 seconds) = 0.0004
                          Correct.

                          Comment


                          • #14
                            Run all these tests configured with LLVM-Clang 3.5.

                            Comment


                            • #15
                              Originally posted by smitty3268 View Post
                              Yeah, i remember that and thinking "how the hell does michael think this makes any sense at all"?
                              Things make more sense when you wrote them. I have to wait for a while between writing and proof reading if I want to have a chance of finding all the stupids. Today I was reading through an analysis I wrote a few months ago and it was glaringly obvious that I had missed something important. At the time... everything looked fine.

                              AFAIK the reason newspapers & magazines have editors is not so much that writers can't write, but that writers can't find problems in articles *they* wrote while the writing is still fresh in their heads... so you either need to slow things way down or use a different head for review. That's why even a quick & dirty peer review works so well.
                              Last edited by bridgman; 08-14-2014, 01:49 AM.

                              Comment

                              Working...
                              X