From f20e479c59d7a1e4cd9dfaaed886821f7523ebd6 Mon Sep 17 00:00:00 2001
From: Josh Lay <me@jlay.io>
Date: Thu, 5 May 2022 21:48:35 -0500
Subject: [PATCH] provide a multiplier to reduce AMDGPU power cap

---
 playbook.yml            | 6 +++++-
 templates/tuned.conf.j2 | 3 ++-
 2 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/playbook.yml b/playbook.yml
index e66522e..bf33ae6 100644
--- a/playbook.yml
+++ b/playbook.yml
@@ -2,6 +2,10 @@
 - hosts: localhost
   become: yes
   vars:
+    power_max_multi: 0.90 # should not exceed 1.0, must be a float. note: driver seems to do its own rounding
+    power_max: "{{ power_max_b64['content'] | b64decode }}"
+    power_cap_float: "{{ power_max |float * power_max_multi }}"
+    power_cap: "{{ power_cap_float |int }}"
     card: card0 # default to card0
     base_profiles: # standard tuned profiles available on Fedora, should dynamically discover?
       - balanced
@@ -37,7 +41,7 @@
     - name: get max power capability for {{ card }}
       slurp:
         src: "{{ hwmon.files.0.path }}"
-      register: power_max
+      register: power_max_b64
     - name: ensure tuned is installed
       package:
         name: tuned
diff --git a/templates/tuned.conf.j2 b/templates/tuned.conf.j2
index 4ff034e..23655fa 100644
--- a/templates/tuned.conf.j2
+++ b/templates/tuned.conf.j2
@@ -5,7 +5,8 @@ summary={{ item.1 }} + TCP/RAID tweaks + AMDGPU pp_power_profile_mode = {{ item.
 [sysfs]
 /sys/class/drm/{{ card }}/device/power_dpm_force_performance_level = manual
 /sys/class/drm/{{ card }}/device/pp_power_profile_mode = {{ item.0.value }}
-/sys/class/drm/{{ card }}/device/hwmon/hwmon9/power1_cap = {{ power_max['content'] | b64decode }}
+# limit to {{ power_max_multi * 100.0 |int }}% of the max power cap, nearest (higher) whole number
+/sys/class/drm/{{ card }}/device/hwmon/hwmon9/power1_cap = {{ power_cap }}
 
 [sysctl]
 net.core.default_qdisc=fq