Thanks for sharing practically my thoughts )I'm also observing all of this, here and at the Luxx. And I honestly don't know what all these details. If I understood what all these "More" do and cause, then I would include it in my tutorial, or write it again here with pictures. But mentally it all left me behind at some point in the spring. And no matter how stupid, or evil, or whatever it sounds, I don't expect that a good manual will be created elsewhere - that was the original motivation for my tutorial, to make everything about "Navi 2X vs. MPT" understandable get together.
In my understanding, Hellm "simply" collects everything that is in the AMD drivers and gives us that as controllable functions in the MPT and Veii was started to reveal the last secrets of the drivers including the bios of the cards. The current intermediate status is this More page, which in my opinion no longer helps normal OCers, but only for the very curious - because all the efforts have so far not resulted in big new OC high scores. Veii is still (trying) to find something, but in the last few weeks his communications have not been too enthusiastic either.
Thinking of the Pareto principle, I think we're through. The last 1-3% performance now has to be scraped out with 200-300% effort, and it is uncertain whether it can even work.
It just so happened for the unfortunate reasons that I get my hands on Navi2x basically at the end of its product cycle, and thus being curious enough to dig into undocumented features I personally feel it's like "too-lazy-cause-too-late" .
Anyway, some points to complete my thoughts:
- most (95%) of the pp table entries for the gfx10 are no secret and has pretty clear designation (at least @me and anyone who tried to search for the info and not just ask for it in forums)
- fully agree without knowing (I mean literally, not just a guesswork) what a certain parameter is responsible for the best case outcome is a waste of time
- if I was the mpt developer i'd start from the SMU metrics and SVI2 telemetry monitoring soft first, to be able to cover as much of dynamically changing operating points or sensors as the firmware provides. The list of available metrics:
CurrClock[GFXCLK]
CurrClock[SOCCLK]
CurrClock[UCLK]
CurrClock[FCLK]
CurrClock[DCLK0]
CurrClock[VCLK1]
CurrClock[DCLK0]
CurrClock[VCLK1]
CurrClock[DCEFCLK]
CurrClock[DISPCLK]
CurrClock [PIXCLK]
CurrClock[PHYCLK]
CurrClock[DTBCLK]
AverageGfxclkFrequencyPreDs
AverageGfxclkFrequencyPostDs
AverageFclkFrequencyPreDs
AverageFclkFrequencyPostDs
AverageUclkFrequencyPreDs
AverageUclkFrequencyPostDs
AverageGfxActivity
AverageUclkActivity
CurrSocVoltageOffset
CurrGfxVoltageOffset
CurrMemVidOffset
AverageSocketPower
TemperatureEdge
TemperatureHotspot
TemperatureMem
TemperatureVrGfx
TemperatureVrMem0
TemperatureVrMem1
TemperatureVrSoc
TemperatureLiquid0
TemperatureLiquid1
TemperaturePlx
AccCnt
ThrottlingPercentage[TEMP_EDGE]
ThrottlingPercentage[TEMP_HOTSPOT]
ThrottlingPercentage[TEMP_MEM]
ThrottlingPercentage[TEMP_VR_GFX]
ThrottlingPercentage[TEMP_VR_MEM0]
ThrottlingPercentage[TEMP_VR_MEM1]
ThrottlingPercentage[TEMP_VR_SOC]
ThrottlingPercentage[TEMP_LIQUID0]
ThrottlingPercentage[TEMP_LIQUID1]
ThrottlingPercentage[TEMP_PLX]
ThrottlingPercentage[TDC_GFX]
ThrottlingPercentage[TDC_SOC]
ThrottlingPercentage[PPT0]
ThrottlingPercentage[PPT1]
ThrottlingPercentage[PPT2]
ThrottlingPercentage[PPT3]
ThrottlingPercentage[FIT]
ThrottlingPercentage[PPM]
ThrottlingPercentage[APCC]
LinkDpmLevel
CurrFanPwm
CurrFanSpeed
EnergyAccumulator
AverageVclk0Frequency
AverageDclk0Frequency
AverageVclk1Frequency
AverageDclk1Frequency
VcnUsagePercentage0
VcnUsagePercentage1
PcieRate
PcieWidth
AverageGfxclkFrequencyTarget
ApuSTAPMSmartShiftLimit
AverageApuSocketPower
ApuSTAPMLimit - in addition to (and followed from) the above, I'd base the tuning utility partly on a direct SMC calls which are well described in a amdgpu driver sources, due to its direct nature and therefore no need to restart driver to show and observe the effect immediately. For this, we'd probably need to just modify the existing kernel driver (inpoutx64) for the PCI device r/w access and memcopy of the tables content to/from user/kernel space, but I'm not sure of that. Of course not excluding "legacy" tinkering with the PowerPLay and OverDrive (and a bunch of other) tables modification to bootup.
Zuletzt bearbeitet
: