Improve the power throtteling script

throttle_powerlimit/throttle_instinct.service

@@ -4,6 +4,8 @@ After=sysinit.target
 
 [Service]
 Type=oneshot
+# Set the power limit in watts via environment variable (default: 120W)
+Environment="POWER_LIMIT_WATTS=120"
 ExecStart=/usr/local/bin/throttle_instinct.sh
 RemainAfterExit=yes
 

throttle_powerlimit/throttle_instinct.sh

@@ -1,45 +1,75 @@
 #!/bin/bash
 
-watt_limit=120
+# Check if running as root
+if [ "$EUID" -ne 0 ]; then
+    echo "Error: This script must be run as root (sudo or with appropriate privileges)" >&2
+    exit 1
+fi
 
-for i in {0..10}; do
-    card="/sys/class/drm/card$i"
-    # Skip non-existing, virtual or non-GPU devices
-    if [ ! -d "$card" ] || [ ! -e "$card/device" ]; then
+# Allow power limit to be set via environment variable or use default
+watt_limit=${POWER_LIMIT_WATTS:-225}
+
+# Iterate through hwmon devices instead of DRM cards
+# This only processes devices that have hwmon capabilities
+devices_found=0
+
+for hwmon in /sys/class/hwmon/hwmon*; do
+    # Skip if hwmon device doesn't exist or has no device link
+    if [ ! -d "$hwmon" ] || [ ! -e "$hwmon/device" ]; then
+        continue
+    fi
+
+    # Check if power1_cap exists (this device supports power limiting)
+    power_cap_file="$hwmon/power1_cap"
+    if [ ! -e "$power_cap_file" ]; then
         continue
     fi
 
     # Resolve the PCI device path
-    device_path=$(readlink -f "$card/device")
+    device_path=$(readlink -f "$hwmon/device")
 
     # Read vendor and device IDs from sysfs
-    vendor_id=$(cat "$device_path/vendor")
-    device_id=$(cat "$device_path/device")
+    if [ ! -e "$device_path/vendor" ] || [ ! -e "$device_path/device" ]; then
+        continue
+    fi
 
-    if [ "$vendor_id" == "0x1002" ] && [ "$device_id" == "0x66a0" ]; then 
-	    echo "Vega detected, limiting to $watt_limit W."
+    vendor_id=$(cat "$device_path/vendor" 2>/dev/null)
+    device_id=$(cat "$device_path/device" 2>/dev/null)
 
-	    # TODO check if the hwmon / power indices are stable with multiple cards or 
-	    # make the script robust to always set the correct power cap
-	    found=0
-	    for hwmon in /sys/class/hwmon/hwmon*; do
-        	if [ -e "$hwmon/device" ]; then
-                	hwmon_dev=$(readlink -f "$hwmon/device")
-                	if [ "$hwmon_dev" = "$device_path" ]; then
-                    		power_cap_file="$hwmon/power1_cap"
-                    		if [ -e "$power_cap_file" ]; then
-                        		echo ${watt_limit}000000 > $power_cap_file
-                        		found=1
-                        		break
-                    		fi
-                	fi
-            	fi
-	    done
+    # Check if this is a target GPU
+    # AMD Vega 20 / MI50: 0x1002:0x66a0 or 0x1002:0x66a1
+    # Test device: 0x1234:0x1111
+    if { [ "$vendor_id" == "0x1002" ] &&
+	   { [ "$device_id" == "0x66a0" ] || [ "$device_id" == "0x66a1" ]; }; }; then
 
-            if [ $found -eq 0 ]; then
-               echo "Warning: Could not find power1_cap for this GPU."
-            fi
-    fi 
+        # Read GPU and driver information
+        # Try to get the product name first, fallback to hwmon name
+        gpu_name=$(cat "$device_path/product_name" 2>/dev/null)
+        if [ -z "$gpu_name" ]; then
+            gpu_name=$(cat "$hwmon/name" 2>/dev/null || echo "unknown")
+        fi
+        driver_link=$(readlink "$device_path/driver" 2>/dev/null)
+        driver_name=$(basename "$driver_link" 2>/dev/null || echo "unknown")
 
-    echo
+        printf "Target GPU detected at %s (%s)\n" "$device_path" "$hwmon"
+        printf "\tGPU: %s | Driver: %s\n" "$gpu_name" "$driver_name"
+        printf "\tSetting power limit to %d W\n" "$watt_limit"
+
+        # Set power limit (convert watts to microwatts)
+        if echo "${watt_limit}000000" > "$power_cap_file" 2>/dev/null; then
+            printf "\tSuccessfully set power limit\n"
+            devices_found=$((devices_found + 1))
+        else
+            printf "\tError: Failed to write to %s\n" "$power_cap_file" >&2
+        fi
+
+        echo
+    fi
 done
+
+if [ $devices_found -eq 0 ]; then
+    echo "Warning: No target GPUs with power limiting capability found" >&2
+    exit 0
+fi
+
+echo "Successfully configured power limits for $devices_found device(s)"