已经将脚本路径、提权指令以及 Systemd 守护进程中的执行路径**全部严格统一为 /usr/local/bin/gpu_fan_control.sh**。同时，将服务名称也顺理成章地对齐为 gpu-fan-control.service，方便日后管理。

直接复制并一次性粘贴到终端执行即可：

1. 生成并赋予核心脚本权限

sudo cat > /usr/local/bin/gpu_fan_control.sh << 'EOF'
#!/bin/bash

# ==========================================
# CPU/GPU 双擎温度联合判定风扇控制脚本 (双向迟滞优化版)
# ==========================================

# ---------------------------------------------------------
# 【模块 1：动态硬件寻址】
# ---------------------------------------------------------
find_hwmon_by_name() {
    local name=$1
    for d in /sys/class/hwmon/hwmon*; do
        if [ -f "$d/name" ] && [ "$(cat "$d/name")" = "$name" ]; then
            echo "$d"
            return 0
        fi
    done
    return 1
}

NCT_DIR=$(find_hwmon_by_name "nct6793")
if [ -z "$NCT_DIR" ]; then
    echo "$(date): FATAL ERROR - Cannot find nct6793 hwmon module." >> /var/log/gpu_fan.log
    exit 1
fi
PWM_PATH="${NCT_DIR}/pwm2"
PWM_EN="${PWM_PATH}_enable"

GPU_DIR=$(find_hwmon_by_name "amdgpu")
if [ -z "$GPU_DIR" ]; then
    echo "$(date): FATAL ERROR - Cannot find amdgpu hwmon module." >> /var/log/gpu_fan.log
    exit 1
fi
if [ -f "${GPU_DIR}/temp2_input" ]; then
    GPU_TEMP_PATH="${GPU_DIR}/temp2_input"
else
    GPU_TEMP_PATH="${GPU_DIR}/temp1_input"
fi

CPU_DIR=$(find_hwmon_by_name "coretemp")
if [ -z "$CPU_DIR" ] || [ ! -f "${CPU_DIR}/temp1_input" ]; then
    echo "$(date): FATAL ERROR - Cannot find CPU coretemp sensor." >> /var/log/gpu_fan.log
    exit 1
fi
CPU_TEMP_PATH="${CPU_DIR}/temp1_input"

# ---------------------------------------------------------
# 【模块 2：用户温控策略配置区】
# ---------------------------------------------------------
GPU_MIN_TEMP=50
GPU_MAX_TEMP=75
CPU_MIN_TEMP=55
CPU_MAX_TEMP=85

CRITICAL_TEMP=90
CRITICAL_PWM=255

FAN_MIN_PWM=40
FAN_MAX_PWM=200

CHECK_INTERVAL=5
PWM_DECAY_STEP=8    # 声学阻尼：降温时，单次最大允许降低的 PWM 步长。避免“转速跳水”
PWM_HYSTERESIS=5    # 迟滞死区：目标 PWM 变化绝对值 <= 5 时无视变化。屏蔽 1~2℃ 的波动杂音
# ---------------------------------------------------------

LAST_PWM=-1
LOOP_COUNT=0

# ---------------------------------------------------------
# 【模块 3：安全兜底与进程退出机制】
# ---------------------------------------------------------
cleanup() {
    [[ -f "${PWM_EN}" ]] && echo 0 > "${PWM_EN}" 2>/dev/null
    echo "$(date): Fan controller stopped, restore fan control to BIOS." >> /var/log/gpu_fan.log
    exit 0
}
trap cleanup SIGTERM SIGINT

if ! echo 1 > "${PWM_EN}" 2>/dev/null; then
    echo "$(date): FATAL ERROR - Open PWM enable failed, resource busy." >> /var/log/gpu_fan.log
    exit 1
fi

echo "$(date): Fan service start success | GPU source:${GPU_TEMP_PATH}" >> /var/log/gpu_fan.log

# ---------------------------------------------------------
# 【模块 4：常驻温控主循环】
# ---------------------------------------------------------
while true; do
    [[ -f "${PWM_EN}" ]] && echo 1 > "${PWM_EN}" 2>/dev/null

    if [ -f "${GPU_TEMP_PATH}" ]; then
        GPU_RAW=$(cat "${GPU_TEMP_PATH}" 2>/dev/null || echo 0)
        GPU_TEMP=$((GPU_RAW / 1000))
    else
        GPU_TEMP=0
    fi

    if [ -f "${CPU_TEMP_PATH}" ]; then
        CPU_RAW=$(cat "${CPU_TEMP_PATH}" 2>/dev/null || echo 0)
        CPU_TEMP=$((CPU_RAW / 1000))
    else
        CPU_TEMP=0
    fi

    if [ "${CPU_TEMP}" -ge "${CRITICAL_TEMP}" ] || [ "${GPU_TEMP}" -ge "${CRITICAL_TEMP}" ]; then
        TARGET_PWM=${CRITICAL_PWM}
        DOMINANT="CRITICAL"
    else
        # [GPU 独立计算]
        if [ "${GPU_TEMP}" -le "${GPU_MIN_TEMP}" ]; then
            PWM_G=${FAN_MIN_PWM}
        elif [ "${GPU_TEMP}" -ge "${GPU_MAX_TEMP}" ]; then
            PWM_G=${FAN_MAX_PWM}
        else
            PWM_G=$(( (GPU_TEMP - GPU_MIN_TEMP) * (FAN_MAX_PWM - FAN_MIN_PWM) / (GPU_MAX_TEMP - GPU_MIN_TEMP) + FAN_MIN_PWM ))
        fi

        # [CPU 独立计算]
        if [ "${CPU_TEMP}" -le "${CPU_MIN_TEMP}" ]; then
            PWM_C=${FAN_MIN_PWM}
        elif [ "${CPU_TEMP}" -ge "${CPU_MAX_TEMP}" ]; then
            PWM_C=${FAN_MAX_PWM}
        else
            PWM_C=$(( (CPU_TEMP - CPU_MIN_TEMP) * (FAN_MAX_PWM - FAN_MIN_PWM) / (CPU_MAX_TEMP - CPU_MIN_TEMP) + FAN_MIN_PWM ))
        fi

        # [取高者原则]
        if [ "${PWM_C}" -gt "${PWM_G}" ]; then
            TARGET_PWM=${PWM_C}
            DOMINANT="CPU"
        else
            TARGET_PWM=${PWM_G}
            DOMINANT="GPU"
        fi

        # [迟滞死区过滤 Hysteresis / Deadband]
        if [ "${LAST_PWM}" -ne -1 ] && [ "${DOMINANT}" != "CRITICAL" ]; then
            PWM_DIFF=$(( TARGET_PWM - LAST_PWM ))
            [ "${PWM_DIFF}" -lt 0 ] && PWM_DIFF=$(( -PWM_DIFF )) # 纯净 Bash 取绝对值法
            
            # 如果变动幅度在迟滞死区内，强制锁定为上一次的转速
            if [ "${PWM_DIFF}" -le "${PWM_HYSTERESIS}" ]; then
                TARGET_PWM=${LAST_PWM}
            fi
        fi

        # [渐进降温阻尼过滤 Decay]
        if [ "${LAST_PWM}" -ne -1 ] && [ "${TARGET_PWM}" -lt "${LAST_PWM}" ]; then
            CALC_DECAY=$((LAST_PWM - PWM_DECAY_STEP))
            if [ "${CALC_DECAY}" -gt "${TARGET_PWM}" ]; then
                TARGET_PWM=${CALC_DECAY}
            fi
        fi

        # [硬性边界钳位]
        ((TARGET_PWM < FAN_MIN_PWM)) && TARGET_PWM=${FAN_MIN_PWM}
        ((TARGET_PWM > CRITICAL_PWM)) && TARGET_PWM=${CRITICAL_PWM}
    fi

    # [执行硬件指令]
    if [ "${TARGET_PWM}" -ne "${LAST_PWM}" ]; then
        if echo "${TARGET_PWM}" > "${PWM_PATH}" 2>/dev/null; then
            LAST_PWM=${TARGET_PWM}
        else
            echo "$(date): WARN write PWM ${TARGET_PWM} failed" >> /var/log/gpu_fan.log
        fi
    fi

    if [ "${LOOP_COUNT}" -eq 0 ]; then
        echo "$(date '+%Y-%m-%d %H:%M:%S') | CPU:${CPU_TEMP}℃ | GPU:${GPU_TEMP}℃ | Source:${DOMINANT} | PWM:${TARGET_PWM}" >> /var/log/gpu_fan.log
    fi
    
    ((LOOP_COUNT=(LOOP_COUNT+1)%6))

    sleep "${CHECK_INTERVAL}"
done
EOF

sudo chmod +x /usr/local/bin/gpu_fan_control.sh

2. 生成 Systemd 守护进程与日志轮转配置

# 生成 Systemd 服务文件 (路径已精准对齐)
sudo cat > /etc/systemd/system/gpu-fan-control.service << 'EOF'
[Unit]
Description=NCT6793 CPU+GPU Smart Fan Controller
Requires=lm-sensors.service
After=lm-sensors.service multi-user.target

[Service]
Type=simple
ExecStart=/bin/bash /usr/local/bin/gpu_fan_control.sh
Restart=on-failure
RestartSec=3
StandardOutput=null
StandardError=journal

[Install]
WantedBy=multi-user.target
EOF

# 生成 Logrotate 日志清理规则
sudo cat > /etc/logrotate.d/gpu-fan << 'EOF'
/var/log/gpu_fan.log {
    daily
    rotate 7
    compress
    missingok
    copytruncate
    notifempty
}
EOF

3. 应用配置并启动服务

重载并启动全新的规范化服务

sudo systemctl daemon-reload
sudo systemctl enable gpu-fan-control.service
sudo systemctl restart gpu-fan-control.service
sudo systemctl status gpu-fan-control.service

@kenshin 说:

风道被打乱了。整机散热要考虑。风应该从散热片抽出。不是对着吹。

這個是局部風流, 只要機箱的風流方向一開始處理好就不會有問題

吹出跟吹入分別不大

@pangfat 说:

@applejuice 你这是涡轮卡吧？靠得太近了，上面那个吸不到风，散热也就困难，我的PCIe间隔3-slot我都觉得太近。

之前插2slot nvlink 所以被逼这样

但是nvlink遇到问题了
寄回给 淘宝的华南金牌旗舰店，他们寄给供应商
供应商说没问题
但是 华南金牌旗舰店 自己测nvlink也是有问题
还在扯皮

如果再不回来我就要把第2张卡移下一格了

@mark 说:

太牛了 . 加风扇的效果, 还不如 配空调, 我感觉 .

冷氣只是將周邊的空氣降溫, 硬件本體還是需要風扇帶來的氣流進行散熱

@joker_chang 说:

我的解决方案如下：

看你图片的意思，两把小风扇是往外吹风，相当于把里面的热风吸走，是这个原理吧？

1.35其實算高了

我DDR5 6400 也就是跑在1.3附近

而且4條來説很難散熱, 建議用個小風扇吹著

我上一間公司有試過因爲顯卡在機箱内排熱結果把内存給熱到當掉了

@terry 说:

这个内存条散热真的有用吗？会不会是反向效果

這個内存散熱外殼有用, 但是得必須依靠機箱風流或者CPU的風冷塔散才能發揮效果

毕竟机器就放在办公桌下（蹭公司的电费），涡轮太吵了。

z40? 我也用一样机箱 但是侧板装起来
前面3个风扇抽风
后面上面一个风扇排风
里面两个风扇对准GPU

但是我限制功率245w

烤机5分钟
gpu - 70c(fan 70%) /74c (fan 92%)
vram - 82c/84c
cpu - 47c

我感觉我加了那么多风扇 cpu 收益最多