AMD P-State Preferred Core Handling Being Enabled For Linux

[F.Ultra]

Here's two other tips:

Despite what all reviewers say: don't overclock memory past the point where it sends more than 1,25V to the SoC. As the chip will degrade beyond that point. The safest bet for Zen 2 and Zen 3 remains 3200Mhz.

The default AMDGPU and Mesa interaction within Linux also by default set the GPU clock way too high way too fast for too long, even while idleing GNOME/KDE usually the clock speed is set to max forever: wasting energy and decreasing the longevity of the chip. Look into tuning the AMDGPU Power Profile Mode through the /sys interface to resolve this. nvtop and radeontop can be used to monitor clocks.

On Gnome 44.3 with a rtx7900xtx and 3440x1440@240Hz radeontop gives a memclock of 460Mhz and a shaderclock of 40Mhz, draws about 35w of power. This out of the box settings on Ubuntu 23.04

[emansom]

MSI Radeon RX 580 Gaming X 8GB

On Gnome 44.3 with a rtx7900xtx and 3440x1440@240Hz radeontop gives a memclock of 460Mhz and a shaderclock of 40Mhz, draws about 35w of power. This out of the box settings on Ubuntu 23.04

Then it's more of a select issue on certain cards and/or generations, I presume. Here on an MSI Radeon RX 580 Gaming X 8 GB (Polaris) it defaults to the PowerMode profile of 3D_FULL_SCREEN on boot, which makes it ramp up to the highest core and memory clocks forever with the slightest GPU load (idle desktop with Firefox open is sufficient) when not configuring and activating an override custom PowerMode profile.

[HD7950]

Is it possible now to change core ranking manually? My stupid Ryzen CPU has marked as the 3 best cores the 3 worst cores. For that reason i had to disable cppc preferred cores in bios.

[geearf]

On Gnome 44.3 with a rtx7900xtx and 3440x1440@240Hz radeontop gives a memclock of 460Mhz and a shaderclock of 40Mhz, draws about 35w of power. This out of the box settings on Ubuntu 23.04

Just FYI, there's also amdgpu_top, I switched to that one recently but truthfully I did not investigate much.

[Anux]

Despite what all reviewers say: don't overclock memory past the point where it sends more than 1,25V to the SoC. As the chip will degrade beyond that point. The safest bet for Zen 2 and Zen 3 remains 3200Mhz.

Ha thx, I already wondered why I got my 2400 memory up to 3200 with just +0,05 V but anything higher wouldn't even post despite going up to 1,5 V. But generally I'm not interested in OC because more power means more noise and I want silent operation.
With those 3 power levels on the other hand you can perfectly tune the APU for your cooling solution, in standard mode it would constantly draw 88 W package power and my fans started to become hearable after a few minutes (atleast on these hot summer days). Now I have short term power at 100 W midterm at 85 and long at 65 and my fans are not noticeable even if I encode over night and still full performance on short tasks.

The default AMDGPU and Mesa interaction within Linux also by default set the GPU clock way too high way too fast for too long, even while idleing

That must be a bug with your distro? I'm on arch and idle clocks stay super low, package power is at 4 W idle which indicates everything saves power like it should.

Also the iGPU is extremly efficient and only takes 5 - 7 W with furmark!

Next I will try undervolting, can you recommend one of these tools? amdctl zenstates ryzen_smu

[emansom]

...
Next I will try undervolting, can you recommend one of these tools? amdctl zenstates ryzen_smu

ryzen_smu is pretty great for monitoring!

Undervolting I wouldn't recommend, too much troubleshooting for not much uplift. Lowering PPT, EDC, TDC and reducing platform thermal throttle limit (to a much more sane 60C for example) within the UEFI are much more effective to get better performance-per-watt.

[HerrLange]

Any Idea how this behaves on a 7900x3d or 7950x3d with one CCD having the cache and high power efficiency while the other CCD has the high frequency high power cores?

[sobrus]

Here's two other tips:

Despite what all reviewers say: don't overclock memory past the point where it sends more than 1,25V to the SoC. As the chip will degrade beyond that point. The safest bet for Zen 2 and Zen 3 remains 3200Mhz.

You can set lower SoC voltage, just need to disable SoC Uncore OC mode deeper in the bios (it is enabled automatically when enabling XMP on my Gigabyte motherboard).
So, ideally, overclock your RAM without enabling XMP - at least if you have Gigabyte mobo.
My 5950x runs 128GB RAM at 3466Mhz (64GB worked fine at 3600Mhz), SoC voltage is 0.950V. Cheap Viper Steel sticks.

Undervolting doesn't give much (if any due to clock stretching) performance uplift, but far better temperatures.
Mine runs at -0.048V voltage offset + custom per core curve close to -30 for CCD1 and around -25 for CCD2 (worst core is at -19).
Air cooled at PPT 120W, 75C temperature limit and 4800Mhz frequency limit.

Despite tight constraints and CPU fans starting only after reaching 50C (otherwise passively cooled) it doesn't seem to affect Geekbench or other scores at all

https://browser.geekbench.com/v6/cpu/2135726

F.Ultra
Both my RX570 and RX6800XT draw around 7W when idling with dual 1440p 60Hz screens.

[emansom]

Undervolting doesn't give much (if any due to clock stretching) performance uplift, but far better temperatures.
Mine runs at -0.048V voltage offset + custom per core curve close to -30 for CCD1 and around -25 for CCD2 (worst core is at -19).
Air cooled at PPT 120W, 75C temperature limit and 4800Mhz frequency limit.

Again, look into PPT, EDC and TDC for better thermals. PBO curve offset is not the place.

[sobrus]

Again, look into PPT, EDC and TDC for better thermals. PBO curve offset is not the place.

This is not true. Ryzen can hit 90C temperature limit easily by pushing 15W+ through few individual cores (15W+ per each core), long before hitting any of these limits.
They will only work for all-core loads.

The correct way of managing thermals (other than undervolting) is actually setting a temperature limit.
This way the CPU will manage frequency and per-core wattage (both current and voltage) according to actual temperature, current leakage and available cooling capacity.