Overclock.net › Forums › AMD › AMD CPUs › cpufrequtils: the Linux answer to P5 state throttling (Kaveri)?
New Posts  All Forums:Forum Nav:

cpufrequtils: the Linux answer to P5 state throttling (Kaveri)? - Page 2

post #11 of 15
Thread Starter 
FM2+ chips are still 15h, so a utility that h4x0rs p-states for Bulldozer/Piledriver may also work for Kaveri. Or it may not. I haven't had a chance to compile it or try it myself yet.

If you could butt heads with the guy who took over amdmsrtweaker:

https://github.com/mpollice/AmdMsrTweaker

you might get somewhere with it

I agree that having functionality like amdmsrtweaker under teh Lun1x would be pretty sweet. Linux is the only OS that fully supports HSA right now, so the thought of the CPU getting throttled needlessly into oblivion every time you send a workload to the GPU is not particularly appealing. HSA isn't really a thing for most folks, but for the few that use it for their own purposes, tools to affect p-states could come in mighty handy.

As a cautionary note, according to my own experiences with amdmsrtweaker, you won't get full satisfaction out of stopping the throttling unless you disable most/all of the power saving settings (APM, etc) in the UEFI/BIOS.
post #12 of 15
Quote:
Originally Posted by mdocod View Post

Everyone always says these APU's are going to be iGPU bound in gaming so the CPU clock speed doesn't matter, but in my experience performance originates with the CPU in gaming
That was my experience too. I mean with a base clock of 3.7 and turbo of 4, dropping to 3 is a pretty significant hit.
Summit Ridge
(16 items)
 
ASUS R510DP
(8 items)
 
 
CPUMotherboardGraphicsRAM
AMD Ryzen 1600X ASRock Fatal1ty AB350 Gaming K4 HIS Radeon HD7870 IceQ 2x8GB G.Skill TridentZ 3200 CL16 
Hard DriveHard DriveHard DriveHard Drive
Plextor M6S Plus 256GB SSD Toshiba X300 6TB Toshiba X300 6TB Toshiba 2TB 
CoolingOSMonitorMonitor
Deepcool BETA 400 ST Windows 8.1 Pro x64 HP S2031 20" Samsung SyncMaster 932BW 19" 
PowerCaseMouseMouse Pad
Seasonic S12G 750w Lian Li full tower Logitech MX310 SteelSeries 4HD 
CPUMotherboardGraphicsRAM
A10-5750m 3.5Ghz ASUStek A75M FCH HD 8650G + HD 8670M dual graphics 2x4GB Samsung 1600Mhz 
Hard DriveOptical DriveOSMonitor
Toshiba Q300 SSD 120GB Panasonic CD/DVD Windows 8.1 Pro 15.6" 1920x1080 
  hide details  
Reply
Summit Ridge
(16 items)
 
ASUS R510DP
(8 items)
 
 
CPUMotherboardGraphicsRAM
AMD Ryzen 1600X ASRock Fatal1ty AB350 Gaming K4 HIS Radeon HD7870 IceQ 2x8GB G.Skill TridentZ 3200 CL16 
Hard DriveHard DriveHard DriveHard Drive
Plextor M6S Plus 256GB SSD Toshiba X300 6TB Toshiba X300 6TB Toshiba 2TB 
CoolingOSMonitorMonitor
Deepcool BETA 400 ST Windows 8.1 Pro x64 HP S2031 20" Samsung SyncMaster 932BW 19" 
PowerCaseMouseMouse Pad
Seasonic S12G 750w Lian Li full tower Logitech MX310 SteelSeries 4HD 
CPUMotherboardGraphicsRAM
A10-5750m 3.5Ghz ASUStek A75M FCH HD 8650G + HD 8670M dual graphics 2x4GB Samsung 1600Mhz 
Hard DriveOptical DriveOSMonitor
Toshiba Q300 SSD 120GB Panasonic CD/DVD Windows 8.1 Pro 15.6" 1920x1080 
  hide details  
Reply
post #13 of 15
Quote:
Originally Posted by 7850K View Post

That was my experience too. I mean with a base clock of 3.7 and turbo of 4, dropping to 3 is a pretty significant hit.

Let me quote The Stilt on another thread:
Quote:
They do.

However like I said earlier the power management is very fast and the performance penalty from the power balancing actions is neglectable.
As long as the application doesn´t fully stress both of the domains simultaneously (which basically never happens in normal workloads) there is no real performance difference.

The difference is 3DMarks (11, Fire Strike, Skydiver) is less than 1% between the default (GeApm On) and the workaround (GeApm Off) conditions.
Similar differences can be observed in games.

If both of the domains are fully loaded then the power management will not be able to balance the available resources properly.
This does happen is applications such as LuxMark (combined mode, CPU + GPU) however I´m yet to find a single real workload or game which would show a similar difference. Running a CPU stress test (e.g. Prime95) and GPU stress test (e.g. Furmark) at the same time would do the same but really is that relevant?
Quote:
Originally Posted by The Stilt View Post

The 3GHz throttling in Kaveri doesn´t have the effect it appears to have based only on the difference in frequency.

If the CPU domain requires greater power envelope while the CPU throttles to 3GHz frequency the power management will burst the CPU frequency back up to meet the new requirements. Steamroller can change it´s frequency state 500 000 times per second (minus the "dead time" from Pll release). This is the reason why you don´t see any performance difference in most applications or benchmarks between the stock condition and "throttling disabled" condition. The only benchmark I´ve personally seen the difference in is Luxmark (in CPU + GPU) mode as it fully utilizes the CPU but also requires CPU processing power to feed the GPU.

I did my own testing on Fallout 4, Metro 2033, Stalker: CoP and a lot of other games/benchmarks (Lost Planet 2, RE 5, AvP)
I can attest that the performance impact in gaming is negligible when using the iGPU: There is no need to modify the P5 state. If needed/wanted, i can post the results (7650k)

However, i encourage the OP to continue his work. thumb.gif
post #14 of 15
Satisfied?
post #15 of 15
Thread Starter 
Quote:
Originally Posted by Moumoule View Post

Let me quote The Stilt on another thread:
I did my own testing on Fallout 4, Metro 2033, Stalker: CoP and a lot of other games/benchmarks (Lost Planet 2, RE 5, AvP)
I can attest that the performance impact in gaming is negligible when using the iGPU: There is no need to modify the P5 state. If needed/wanted, i can post the results (7650k)

However, i encourage the OP to continue his work. thumb.gif

Bear in mind that you are covering gaming tests. The OP is concerned with defeating the throttling behavior under Linux, where gaming is . . . shall we say, less prominent. If you load up the iGPU with a compute function like . . . GPUPi (which does not, to my knowledge, exist under Linux, but it could) and then try to do something CPU-intensive underneath it, the CPU will lock itself down to the lower clockspeed and kill performance for the CPU task.

For example, in Windows, try running GPUPi + one instance of SuperPi at the same time. The SuperPi score will tank.

So, it's all a matter of what it is you are doing with your Kaveri.
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: AMD CPUs
Overclock.net › Forums › AMD › AMD CPUs › cpufrequtils: the Linux answer to P5 state throttling (Kaveri)?