Personal computing discussed
Moderators: renee, mac_h8r1, Nemesis
synthtel2 wrote:Updated with the R7 1700 I just got. I can't get cpufreq to give me much granularity on that, unfortunately. I've got three modes - 1.55 = 830mV, 2.7 = 910mV, and "3.0" = 1060mV. The so-called 3.0 mode is actually more like 3.15-3.2 AFAICT. Single-core loading gets it to about 3.6 and 1310mV in the testing I've done. I've never seen it reach 3.7. When idle (background load only), it can be anywhere from 670mV to 1310mV (it's very turbo-happy).
synthtel2 wrote:I think most of the Zen OC data I've seen is 1.2-1.3V for 3.7 GHz, so accounting for the extra margin the factory puts on it, 1.31 for 3.6 seems high-ish but close enough to the normal range. My data versus The Stilt's data definitely looks weird at low power, though. That mine goes as low as 670mV when actually idle makes me think some specific part wants more voltage.
Yours is probably more representative of Intel high-performance bins. The two I've got here are both obviously not showing Intel at their best. If you'd be interested in capturing voltage at 900 MHz, ~2.5 GHz, and ~3.0 GHz, I'd definitely like to know how that line trends.
bfg-9000 wrote:Does anyone else find 1.4v on 14nm disturbing? Back at 45nm we had a never-exceed voltage of 1.45v, and at 130nm (nearly 10x larger) we had Sudden Northwood Death Syndrome above 1.7v...
Noinoi wrote:I'll try to see what I can do.
synthtel2 wrote:Noinoi wrote:I'll try to see what I can do.
Thanks!
synthtel2 wrote:Yeah, I've just been using stock curves. OC/UV curves are probably more accurate, but figuring out the limits of stability at each frequency is way too much work. In Windows, I just used power options -> advanced plan settings (or whatever it was called, not on Windows to check) to adjust. That's only annoying because it's a percent instead of an absolute speed and pressing enter closes the window instead of making the new value take effect, necessitating mouse use.
MHz Core1 Core2 Core3 Core4 Average
800 0.702 0.707 0.705 0.703 0.704
900 0.717 0.722 0.720 0.723 0.721
1000 0.727 0.727 0.730 0.727 0.728
1100 0.741 0.746 0.744 0.742 0.743
1200 0.751 0.751 0.749 0.752 0.751
1300 0.761 0.761 0.759 0.761 0.761
1400 0.776 0.776 0.779 0.776 0.777
1500 0.785 0.781 0.784 0.786 0.784
1600 #DIV/0!
1700 0.805 0.805 0.808 0.805 0.806
1800 #DIV/0!
1900 0.829 0.830 0.828 0.830 0.829
2000 0.839 0.840 0.843 0.839 0.840
2100 #DIV/0!
2200 0.863 0.859 0.862 0.864 0.862
2300 #DIV/0!
2400 0.883 0.884 0.887 0.883 0.884
2500 0.897 0.899 0.897 0.898 0.898
2600 0.907 0.904 0.907 0.907 0.906
2700 #DIV/0!
2800 0.927 0.928 0.931 0.932 0.930
2900 0.946 0.943 0.941 0.942 0.943
3000 0.956 0.953 0.956 0.956 0.955
3100 0.966 0.963 0.966 0.966 0.965
3200 #DIV/0!
3300 0.990 0.988 0.990 0.990 0.990
3400 #DIV/0!
3500 1.034 1.032 1.034 1.034 1.034
3600 #DIV/0!
3700 1.078 1.076 1.078 1.078 1.078
synthtel2 wrote:@Noinoi, I updated the graph. Thanks again, that looks much more like Haswell is probably supposed to.
The i3-6100 was on Win7, so there's lots of opportunity for software differences. The same thing applies with the gaps, though.
synthtel2 wrote:4790Ks do need a bit of a lower/later curve than yours has, I think, but not by too much. If yours stayed at that last slope until 4.4 (which it probably doesn't), it would do a 4790K's 4.4 turbo at 1.23V.
synthtel2 wrote:Hmm, I was going to suggest a non-AVX load for less heat generation, as I was only using prime95 because it was handy, but now that you mention it Haswells that support AVX (not the G3258) can run different voltages when it's in use. It probably isn't a huge effect on the shape, though, and a lot of Haswells don't seem to have that issue at all (like my 4690K back when it worked). I'll label your 4590 as an AVX load next time I update the graph - it's probably not critical whether it's on or off for anything here, but still good to track.
MHz Core1 Core2 Core3 Core4 Average
800 0.675 0.678 0.674 0.677 0.676
900 0.689 0.692 0.684 0.691 0.689
1000 0.703 0.707 0.698 0.706 0.704
1100 0.718 0.716 0.713 0.720 0.717
1200 0.732 0.731 0.727 0.730 0.730
1300 0.742 0.745 0.742 0.744 0.743
1400 0.757 0.760 0.757 0.758 0.758
1500 0.771 0.774 0.766 0.773 0.771
1600 0.786 0.784 0.781 0.782 0.783
1700 0.795 0.798 0.796 0.797 0.797
1800 0.810 0.808 0.805 0.811 0.809
1900 0.824 0.822 0.820 0.825 0.823
2000 0.839 0.836 0.834 0.840 0.837
2100 0.848 0.851 0.849 0.849 0.849
2200 0.863 0.860 0.859 0.864 0.862
2300 0.877 0.875 0.873 0.878 0.876
2400 0.891 0.889 0.893 0.892 0.891
2500 0.920 0.918 0.917 0.921 0.919
2600 #DIV/0!
2700 #DIV/0!
2800 #DIV/0!
2900 #DIV/0!
3000 1.026 1.053 1.024 1.050 1.038
synthtel2 wrote:@Noinoi, I updated the graph. Thanks again, that looks much more like Haswell is probably supposed to.
The i3-6100 was on Win7, so there's lots of opportunity for software differences. The same thing applies with the gaps, though.
@Chrispy_ (and thanks DPete27), I don't see any voltages there. It is still an interesting read on the power limits, though.
synthtel2 wrote:Nice! Graph is updated. Yeah, that's pretty different. Considering the corner is probably actually somewhere after 2.4 (resolution issues), that's probably a more GloFo-like corner steepness, and that really does look like a lower threshold (even if it isn't on threshold at 800 MHz). As of Haswell, Intel does still have a different mobile process?
Maybe encoding something or gaming would be good for decreasing heat load? Prime95 on only one core at a time (like I did to get the full turbo on my 1700) could also be useful, and shouldn't shift any voltages by much.
MHz Volts
800 0.677
900 0.691
1000 0.706
1100 0.720
1200 0.730
1300 0.744
1400 0.758
1500 0.773
1600 0.782
1700 0.797
1800 0.811
1900 0.825
2000 0.840
2100 0.849
2200 0.864
2300 0.878
2400 0.892
2500 0.921
2600 0.950
2700 0.974
2800 0.998
2900 1.021
3000 1.050