AMD Zen Thermal/Power Reporting Improvements Could Hit Linux 5.6 But More Testing Needed

[phoronix]

Phoronix: AMD Zen Thermal/Power Reporting Improvements Could Hit Linux 5.6 But More Testing Needed

Last week I eagerly reported on Ryzen CPUs on Linux finally seeing CCD temperatures and current/voltage reporting thanks to new patches to the k10temp driver by Google's Guenter Roeck who oversees the kernel's hardware monitoring "HWMON" subsystem. The patches seem to be working well and are tentatively queued in hwmon-next, but more testing is still needed...

http://www.phoronix.com/scan.php?page=news_item&px=K10temp-Could-Improve-5.6

[birdie]

I have tested the k10temp patches on a number of AMD Zen/Zen+/Zen2 systems so far and have not run into any problems.

By testing Guenter Roeck meant not just verifying that you get some info, Michael

What he really meant is that you needed to reboot into Windows (where you can run either HWiNFO64 or Ryzen Master) and check whether the values you're getting across OS'es match. Alternatively, you could achieve the same if you have the equipment (which barely anyone has except for overclockers, repair centers and AMD engineers) to measure your CPU parameters directly.

Otherwise this testing isn't worth a penny.

You don't necessarily need to buy a Windows license: here's a free 90 days Windows for you.

[shmerl]

Can it work through Wine, or it's doing some very low level stuff for it?

[pcwolf]

I am far too technoob to roll my own and apply patches. I can only stand on the sideline and cheer heartily!

[Michael]

By testing Guenter Roeck meant not just verifying that you get some info, Michael

What he really meant is that you needed to reboot into Windows (where you can run either HWiNFO64 or Ryzen Master) and check whether the values you're getting across OS'es match. Alternatively, you could achieve the same if you have the equipment (which barely anyone has except for overclockers, repair centers and AMD engineers) to measure your CPU parameters directly.

Otherwise this testing isn't worth a penny.

You don't necessarily need to buy a Windows license: here's a free 90 days Windows for you.

I have verified to Windows on some systems but even outside of that already working through with him fixes for new Threadripper CPUs and other verification. He is also looking for idle vs. load confirmation and overclock results.

[playenball]

Looks like there is a Version 3 of the patches posted

https://lore.kernel.org/lkml/[email protected]/

[malakudi]

I tested the patch on my Ryzen 3900X. Here is what it reports while running blender benchmark, bmw27
k10temp-pci-00c3
Adapter: PCI adapter
Vcore: +1.23 V
Vsoc: +1.05 V
Tdie: +63.0°C
Tctl: +63.0°C
Tccd1: +60.2°C
Tccd2: +62.8°C
Icore: +61.93 A
Isoc: +15.01 A

Temps and Vcore, Vsoc seem correct. HWINFO64 displays Vcore 1.231V Vsoc 1.050V. Icore and Isoc are probably not correct. HWINFO64 in Windows 10 doesn't display Icore,Isoc but displays CPU Core Power (113.4W) and CPU Core Power (18.3W). If I do the math, 15.01*1,05=15.76W, different from 18.3W and 61.93*1,231=76,23W, different from 113.4W
So I think Ιcore, Isoc are wrong, not only because they are different from what HWINFO64 reports but also because Ryzen 3900X definitely goes above 95W total power while running blender benchmark with Core Performance Boost enabed. It goes up to 142W which is the PPT limit without PBO.

[birdie]

I tested the patch on my Ryzen 3900X. Here is what it reports while running blender benchmark, bmw27
k10temp-pci-00c3
Adapter: PCI adapter
Vcore: +1.23 V
Vsoc: +1.05 V
Tdie: +63.0°C
Tctl: +63.0°C
Tccd1: +60.2°C
Tccd2: +62.8°C
Icore: +61.93 A
Isoc: +15.01 A

Temps and Vcore, Vsoc seem correct. HWINFO64 displays Vcore 1.231V Vsoc 1.050V. Icore and Isoc are probably not correct. HWINFO64 in Windows 10 doesn't display Icore,Isoc but displays CPU Core Power (113.4W) and CPU Core Power (18.3W). If I do the math, 15.01*1,05=15.76W, different from 18.3W and 61.93*1,231=76,23W, different from 113.4W
So I think Ιcore, Isoc are wrong, not only because they are different from what HWINFO64 reports but also because Ryzen 3900X definitely goes above 95W total power while running blender benchmark with Core Performance Boost enabed. It goes up to 142W which is the PPT limit without PBO.

Now we have real tests, thanks!

I am also not convinced by the Icore/Isoc values that I'm getting on my 3700X.

[malakudi]

Did another test, this time running with power limits of a 65W part (PPT=88,TDC=60,EDC=90). Again running blender benchmark, bmw27
k10temp-pci-00c3
Adapter: PCI adapter
Vcore: +1.01 V
Vsoc: +1.05 V
Tdie: +46.1°C
Tctl: +46.1°C
Tccd1: +43.2°C
Tccd2: +46.5°C
Icore: +39.53 A
Isoc: +14.42 A

Temps and Vcore, Vsoc agree with HWINFO64. Vcore is actually reported as 1.012V. CPU Core Power is reported as 59.59W, so we would expect Icore of 58.36A and not 39.53. CPU SOC Power is reported as 18.05W this time, so Isoc should be 17.19A and not 14.42A

Should I forward these data somewhere or they will be seen here from the devs?

[birdie]

Did another test, this time running with power limits of a 65W part (PPT=88,TDC=60,EDC=90). Again running blender benchmark, bmw27
k10temp-pci-00c3
Adapter: PCI adapter
Vcore: +1.01 V
Vsoc: +1.05 V
Tdie: +46.1°C
Tctl: +46.1°C
Tccd1: +43.2°C
Tccd2: +46.5°C
Icore: +39.53 A
Isoc: +14.42 A

Temps and Vcore, Vsoc agree with HWINFO64. Vcore is actually reported as 1.012V. CPU Core Power is reported as 59.59W, so we would expect Icore of 58.36A and not 39.53. CPU SOC Power is reported as 18.05W this time, so Isoc should be 17.19A and not 14.42A

Should I forward these data somewhere or they will be seen here from the devs?

groeck

Any suggestions on how to collaborate on the new features of the new driver?