Major breakthrough for power management

Someone asked me why he only got a few P-States with his Haswell setup (GA-Z87MX-D3H motherboard) when logIPGStyle – Intel Power Gadget style logging – is set to false in AppleIntelCPUPowerManagementInfo.kext/C*/Info.plist The expected output is something like this:

AICPUPMI: CPU P-States [ 8 17 34 35 36 37 38 ]

Three P-States (8, 17 and 34) plus the four turbo states (35-38). I know. I used to say that this is pretty normal, but now I know better. Yes indeed. I finally figured it out, and as a result his setup now gives him a lot more P-States:

AICPUPMI: CPU P-States [ 8 17 (20) 21 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 ]

Yes. This is with logIGPStyle set to false. The trick was to change two settings in the Gigabyte UEFI BIOS

With the Auto settings giving us this:

MSR_PP0_CURRENT_CONFIG…..(0x601) : 0x1F40
MSR_PKG_POWER_LIMIT……..(0x610) : 0xFFD00000EA82
MSR_PP0_POWER_LIMIT……..(0x638) : 0xFFD0
MSR_PP1_POWER_LIMIT……..(0x640) : 0xFFD0

Note: MSR_PP1_POWER_LIMIT is for the IGPU only!

Then we changed the values from Auto to:

And this is the result:

MSR_PP0_CURRENT_CONFIG…..(0x601) : 0x2F8 (760 >> 3) is 95 Amp
MSR_PP0_POWER_LIMIT……..(0x638) : 0x82A0 (672 >> 3) is 84 Watt
MSR_PP1_POWER_LIMIT……..(0x640) : 0x82A0 (672 >> 3) is 84 Watt

The value 3 above is taken from (MSR_RAPL_POWER_UNIT & 0xf) and that equals to power units of 1/8 Watt – hence the right shift to get the proper value.

We also changed some other settings, this to be certain that XCPM is controlling power management, and not the BIOS:

Let me know if this works for your setup as well.

Thanks!

Update

We already knew that MSR_PMG_CST_CONFIG_CONTROL is re-initialised by XCPM, the mach_kernel, but I also found out that the IA32_ENERGY_PERF_BIAS (0x1B0) MSR is set by data that is taken from the resource plist (think Mac-F60DEB81FF30ACF6.plist et all).

Update-2

Ok. This proves that I was right when I said that having more P-States isn’t necessarily better. In fact. It now results in a (somewhat) slower PC, because the Geekbench v3.1.2 (32-bit mode) score is 3624/11298 without the changes in the UEFI BIOS (auto versus 85 Watts and 95 Amps) and 3613/9501 with the changes.

MacBook Pro data:

MSR_RAPL_POWER_UNIT……..(0x606) : 0xA0E03
MSR_PP0_CURRENT_CONFIG…..(0x601) : 0x1814149480000380
MSR_PKG_POWER_LIMIT……..(0x610) : 0x8000815E00DC8118
MSR_PP0_POWER_LIMIT……..(0x638) : 0x0

MacBook Air data:

MSR_RAPL_POWER_UNIT……..(0x606) : 0xA0E03
MSR_PKG_POWER_LIMIT……..(0x610) : 0x4283E800DD8320
MSR_PP0_CURRENT_CONFIG…..(0x601) : 0x4010141400000100
MSR_PP0_POWER_LIMIT……..(0x638) : 0x0

Update-3

Someone over at Tony’s place with a Gigabyte GA-Z87X-UD5H board (F9 bios) ran into another PM related issue – using ‘Auto’ as core ratio, instead of setting it manually to anything but 35 made PM fail for him.

I checked the MSR log of AppleIntelCPUPowerManagementInfo.kext from another user with a similar issue and found this:

MSR_PLATFORM_INFO……….(0xCE) : 0x80838F301002300

Nothing wrong here. Perfectly fine for his i7-4770K. The detected minimum and maximum core ratios are also fine.

CPU Low Frequency Mode………….: 800 MHz
CPU Maximum non-Turbo Frequency….: 3500 MHz
CPU Maximum Turbo Frequency……..: 3900 MHz

But then I spotted this:

MSR_TURBO_RATIO_LIMIT……(0x1AD) : 0x27272727

What is interesting here is that all core turbo ratios are set to 27 (3.9GHz) and thus a single turbo request will result in the top frequency of the processor. Completely ignoring the fact that the first turbo frequency must be set to the core ratio + 1. In this case 36. Otherwise we ran into issue ages ago already. Remember?

I have asked the person to set the first turbo ratio back to 36 and the core ratio to 35 (or auto) to see what this does. I’ll keep you posted if I have anything new to report.

Thanks to Toleda for the heads up 😉

Update-4

Both MSR_PP0_POWER_LIMIT and MSR_PP1_POWER_LIMIT are set to zero (0) on Apple hardware, and using two lines in the boot loader:

wrmsr(MSR_PP0_POWER_LIMIT, 0)
wrmsr(MSR_PP1_POWER_LIMIT, 0)

Didn’t change anything. Which is positive.