Announcement

Collapse
No announcement yet.

Very low floating point math score.

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

  • Very low floating point math score.

    Hi!
    I get a very, very low score on the floating point. (~2450)
    An i3 is get higher results.

    I have an overclocked 6600k (45x102.25) ~4.6GHz.

    No temp or throttling problem.
    This cant have anything to do with my memory, because all the memory test gives me pretty high scores.

    I cannot for my life understand why i get so low score?
    Anyone have any idea?

    Running latest version as of October 11 and latest Windows 10.

  • #2
    If you return the CPU to normal clock speed does it fix the problem.
    If you do several runs is it consistently bad?
    If you look in task manager are all cores being loaded to 100% during the CPU test (this is easier to see if you pick a long test duration from the preferences window)
    Are the other CPU benchmark scores OK?

    Comment


    • #3
      Originally posted by David (PassMark) View Post
      If you return the CPU to normal clock speed does it fix the problem.
      If you do several runs is it consistently bad?
      If you look in task manager are all cores being loaded to 100% during the CPU test (this is easier to see if you pick a long test duration from the preferences window)
      Are the other CPU benchmark scores OK?
      I tried optimized default in bios, i get around 200-300 points lower.
      As far as i can see cores are working 100%.

      Click image for larger version  Name:	passmark.png Views:	3 Size:	51.3 KB ID:	36099

      Comment


      • #4
        If you look in task manager during the Floating point test, can you check that

        1) All cores being loaded to 100% during the CPU test
        2) The CPU clock speed ramps up and stays at a minimum of 3.5Ghz (it won't hit the 3.9Ghz Turbo speed as all cores should be under load).

        The above will be easier to see if you pick a very long test duration from the preferences window and run the Floating point test just by itself from the Menu.
        Do this without any overclocking as it keeps thing simpler.

        Comment


        • #5
          Originally posted by David (PassMark) View Post
          If you look in task manager during the Floating point test, can you check that

          1) All cores being loaded to 100% during the CPU test
          2) The CPU clock speed ramps up and stays at a minimum of 3.5Ghz (it won't hit the 3.9Ghz Turbo speed as all cores should be under load).

          The above will be easier to see if you pick a very long test duration from the preferences window and run the Floating point test just by itself from the Menu.
          Do this without any overclocking as it keeps thing simpler.
          Yes, confirmed all cores 100%. And with my 4.6GHz o/c it hits around 4.5GHz.
          2398 points...

          Comment


          • #6
            Again, would have been better to to this without O/C. I haven't really looked into it, but I suspect you shouldn't be able to get to 4.5Ghz on all cores under full load. Where as a 4.5Ghz turbo single core O/C I could believe.

            I did some data mining in our database. At the moment we have 2028 examples of i5-6600K benchmarks. Of these 45 units (2%) exhibit similar behaviour of bad floating point but otherwise good CPU resutls. From these 45 "bad" units I picked 6 at random for a closer look. In each case they were overclocked units (4.7Ghz turbo in one case). The bad units were tested using various different versions of PerformanceTest that were released over the last couple of years (not just V9).

            I then did the same analysis with the common i7-2600 & i7-2600k. From the ~8000 samples there was zero units that had the same issue.

            My conclusion is that something goes wrong with some of these i5-6600K when they are overclocked, maybe they are susceptible to being burnt out when the voltages or temps are too high. Or maybe there is some internal throttling mechanism that we aren't aware of in this CPU.



            Comment


            • #7
              Originally posted by David (PassMark) View Post
              Again, would have been better to to this without O/C. I haven't really looked into it, but I suspect you shouldn't be able to get to 4.5Ghz on all cores under full load. Where as a 4.5Ghz turbo single core O/C I could believe.

              I did some data mining in our database. At the moment we have 2028 examples of i5-6600K benchmarks. Of these 45 units (2%) exhibit similar behaviour of bad floating point but otherwise good CPU resutls. From these 45 "bad" units I picked 6 at random for a closer look. In each case they were overclocked units (4.7Ghz turbo in one case). The bad units were tested using various different versions of PerformanceTest that were released over the last couple of years (not just V9).

              I then did the same analysis with the common i7-2600 & i7-2600k. From the ~8000 samples there was zero units that had the same issue.

              My conclusion is that something goes wrong with some of these i5-6600K when they are overclocked, maybe they are susceptible to being burnt out when the voltages or temps are too high. Or maybe there is some internal throttling mechanism that we aren't aware of in this CPU.


              Interesting find!
              I did a default run with optimized defaults in bios and the difference in floating point math score is more or less within a margin of error, 200-300 points up/down.
              I will a try a few things here.

              I don't think there is any "hidden" or secret internal throttling or something like that, all other benchmarks and burn-tests works like a charm.
              Aida, Prime95, Intelburn test, Cinebench, realbench, all of these are working fine.

              Comment


              • #8
                Ok, i have now tried different settings in bios aswell as an optimized default.
                Regardless if i run the CPU @stock, @4.2GHz or @4.6GHz, i get pretty much the same score.
                It sits around 2300-2500 points.

                What kind of calculation is this specific test doing? Is there any other software i can check this with?

                Comment


                • #9
                  As the benchmark name of the test implies, floating point maths is being done.

                  The test is really simple & the code hasn't changed for years.

                  Simplified, it looks like this,

                  float *testdata;
                  double *pdtestdata;

                  while (1) {
                  val1_f = testdata[ind];
                  val1_d = pdtestdata[ind++];
                  val2_f = testdata[ind];
                  val2_d = pdtestdata[ind++];

                  j_f=val1_f + val2_f; // Line 1
                  j_f=val1_f - val2_f;
                  j_f=val1_f * val2_f;
                  j_f=val1_f + val2_f;
                  j_f=val1_f - val2_f;
                  j_f=val1_f * val2_f;
                  j_f=val1_f + val2_f;
                  j_f=val1_f - val2_f;
                  j_f=val1_f * val2_f;
                  j_f=val1_f / val2_f; // Line 10
                  <snip>

                  //64-bit Double math added in PT6.0
                  j_d=val1_d + val2_d; // Line 1
                  j_d=val1_d - val2_d;
                  j_d=val1_d * val2_d;
                  j_d=val1_d + val2_d;
                  j_d=val1_d - val2_d;
                  j_d=val1_d * val2_d;
                  j_d=val1_d + val2_d;
                  j_d=val1_d - val2_d;
                  j_d=val1_d * val2_d;
                  j_d=val1_d / val2_d; // Line 10
                  <snip>

                  //If time expired quit loop and calculate operations per second
                  <snip>

                  } //loop end

                  This loop is threaded and runs on all available cores (or the number specified by the user in the preferences window).

                  Comment


                  • #10
                    Posting that code snippet above made me thing of something else to try as an experiment.

                    In the Preferences window change the number of processes to 1, then 2, then 3... etc... and record the result each time.

                    You should see a progression like this.
                    .
                    Floating point test, number of processes Benchmark result, Operations/Sec
                    1 2276
                    2 4342
                    3 6354
                    4 7545
                    5 7847
                    6 7761
                    7 7728
                    8 7382
                    9 7417
                    .
                    This chart was from a 4 core i7-4770. So you see near linear scaling to 4 processes. Then a levelling off and even a slight drop in throughput as more load is added. Which is slightly surprising as hyperthreading should have kicked in and gave some slight improvement from 4 to 8 processes. Hyperthreading always was a bit hit and miss however.

                    Maybe looking at the scaling on your machine would give some insight into the issue.

                    Comment


                    • #11
                      Originally posted by David (PassMark) View Post
                      .
                      This chart was from a 4 core i7-4770. So you see near linear scaling to 4 processes. Then a levelling off and even a slight drop in throughput as more load is added. Which is slightly surprising as hyperthreading should have kicked in and gave some slight improvement from 4 to 8 processes. Hyperthreading always was a bit hit and miss however.

                      Maybe looking at the scaling on your machine would give some insight into the issue.
                      1 core: ~1330
                      2 core: ~2280
                      3 core: ~2430
                      4 core: ~2480

                      Comment


                      • #12
                        So even running 1 core is slow.
                        Still looks like throttling, but that doesn't explain why only the floating point result is effected.
                        So a bit of a puzzle.

                        Other floating point benchmarks are,
                        The Classic Whetstone
                        http://www.roylongbottom.org.uk/whetstone.htm

                        Spec CPU2006
                        https://www.spec.org/cpu2006/CFP2006/

                        But they aren't friendly bits of software to download & run.

                        SANDRA might also have floating point tests. Beware however some of these might use different CPU instructions (SIMD SSE in-particular) so the same circuits in the CPU might not be used as in our test.



                        Comment


                        • #13
                          I have heard few heavily overclocked skylake cpus having problem with AVX and floating point calculation performance coz of its PCU going bad. I'm not sure about the details though. Those were after overclocking competition.

                          Comment


                          • #14
                            I didn't know what the PCU was. PCU = Power Control Unit or Package Control Unit, depending on who you believe. It is a new module inside Intel Skylake CPUs.

                            The best explanation I found was from anandtech.com.
                            "The PCU is essentially a microcontroller that monitors and computes the power requests and consumption portfolio of the separate silicon areas, providing information that can allow parts of the CPU to be power gated, duty cycled, or adjust in frequency and voltage."

                            I also read some that on some motherboards you can disable the PCU, but the side effects are pretty bad. Like the loss of the internal iGPU, the loss of all power control (the CPU never idles as it should) and lower speed from certain instruction mixes. In particular AVX instructions running slower.

                            The bit about duty cycled and frequency adjustment is interesting. So it is plausible this PCU module is now damaged and doing some frequency adjustments that it probably shouldn't be. Going to had to confirm however without swapping out the CPU for a new one.

                            Comment


                            • #15
                              Stumbled on this page while troubleshooting a gaming comp. I had similar case, and it helped me somewhat. Here's what I did in case it helps someone.

                              Win 8.1. 6600k, overclocked to 4.2ghz, no problems in most cases, heavy stuttering in few games. It would go from 40-60fps to single digits in cases where physics was concerned. I had some suspictions it was cpu's fault. Ran passmark, in most cases it was on par with similar systems, but in floating point math, it was very low. Seme results no matter what clock speed I tried. Can't remmember correctly, but it was in 2000 range. Anway, tried it on completely new Win install to rid of potential driver and registry problems. Win 10 on spare SSD (that already had some problems of it's own), and problem was immidately gone. Only tried with stock 3.9ghz. Games without stuttering, passmark score on floating point math: 8837. No idea what exactly was the fault, malware, no win reinstall on new cpu, overclock, drivers... But the problem was gone. Mind also that I'm not completely certain that it was cpu, it may also had something to do with ssd disk that 8.1 were installed on.

                              Comment

                              Working...
                              X