Custom Power Cap for Radeon Pro v620

[Brian Anderson]

Hello,

Thank you for all that you do with your help maintaining the AMD GPU
driver. This issue proposes custom power cap values consistent with
the Radeon Pro v620's SMU that are currently inaccessible due to the
power caps being derived from the VBIOS.

Recently I came across a friction point with the kernel driver for AMD
GPU related to SWSMU. I have a Radeon Pro v620 and after reviewing
community documentation resources I came across a generative AI
assisted report
(https://blivioniag.github.io/gfx1030-wiki/tuning/power.html) that
talked about the power management for the v620, specifically through
the lens of LLM processing with the v620. The v620 does indeed list
it's power_cap_min and power_cap_max as both being 250000000 (250W),
but the report I linked is also correct that the GPU's VBIOS accepts
values lower than that. After following the report and rebuilding the
kernel with changes documented to override the power_cap_min values I
found that the GPU's VBIOS SMU does indeed accept lower power cap
values than 250W.

I understand that this is a difficult issue because it doesn't
represent a true bug in the sense that something is working
incorrectly, as the kernel driver is correctly reading the power cap
min value from the VBIOS. That being said, legitimate functionality is
being obscured here, and there are many use cases that would benefit
from being to communicate with the SMU within it's accepted range.
Would it be possible to include either an override for the v620 or
another means of deriving the power cap in the AMD GPU driver that
could enable setting the SMU within it's accepted range for the Radeon
Pro v620?

Steps to reproduce:

echo 200000000 | tee /sys/class/drm/card{V620 CARD
IDX}/device/hwmon/hwmon{HWMON_VAL}/power1_cap
For instance, I use:

$ echo 200000000 | sudo tee /sys/class/drm/card1/device/hwmon/hwmon2/power1_cap
tee: /sys/class/drm/card1/device/hwmon/hwmon2/power1_cap: Invalid argument
$ cat /sys/class/drm/card1/device/hwmon/hwmon2/power1_cap_min
250000000
$ cat /sys/class/drm/card1/device/hwmon/hwmon2/power1_cap_max
250000000

Kernel version: Linux version 7.2.4-arch1-2
Distribution: Arch Linux
CPU Architecture: x86_64 unknown

Thank you,
Brian Anderson