Power Management

What are the power specifications of the RK182X?

The RK182X is an AI coprocessor, not a general-purpose SoC, so it has no traditional power domains such as VDD_CPU or VDD_GPU. The RK182X is powered through the host-side PCIe/USB/EtherNet interface, and power management focuses on NPU frequency scaling, working-mode switching, sleep/wakeup, and thermal control strategy.

Reference power figures for the RK182X and RK1820 (source: RK182X/RK1820 AI SDK Quick Start V1.0.4, §4.3.1):

ParameterRK182XRK1820
Standby power81 mW—
Typical power—1631 mW
Device power—1681 mW
Standby temperature44°C26°C
NPU DRAM5 GB2.5 GB

Note: The RK182X idle power is 81mW (NPU occupancy 0%) at 44°C. The RK1820 typical power is 1631mW at 26°C. These figures were measured on an RK3588 EVB10 + SODIMM module; actual values vary with hardware design and cooling conditions.

What are the NPU power figures for the RK182X?

NPU current data for LLM inference scenarios (source: RK182X/RK1820 AI SDK Quick Start V1.0.4, §7.4):

ModelNPU frequencyNPU average currentNPU peak currentPrefill Mode
Qwen2.5 3B LLM1000 MHz7.25 A28.37 APERFORMANCE
Qwen2.5 3B LLM1000 MHz7.12 A15.35 AEFFICIENT

Prefill Mode affects the peak current during the Prefill phase of LLM inference:

  • EFFICIENT: NPU frequency in the Prefill phase is half of the Decode phase, with lower peak current (15.35A)
  • PERFORMANCE: NPU runs at the same frequency in Prefill and Decode phases, with higher peak current (28.37A) but lower TTFT (341.5ms vs 582.9ms)

Performance modes and NPU frequency

The RK182X supports adjusting NPU performance levels with the rknn-smi tool; each level maps to a different maximum NPU frequency (source: RK182X/RK1820 AI SDK Quick Start V1.0.4, §7.1):

Performance levelLevel valueSODIMM moduleM.2 module
EFFICIENT0500 MHz400 MHz
NORMAL1850 MHz500 MHz
PERFORMANCE21000 MHz850 MHz
# 查看当前工作模式
rknn-smi info -t work_mode

# 设置为 NORMAL 模式
rknn-smi set -t work_mode -s 1

# 多设备时指定设备
rknn-smi set -t work_mode -s 1 -d 0

How is NPU frequency DVFS configured?

The RK182X's DVFS (Dynamic Voltage and Frequency Scaling) is implemented through dynamic NPU frequency adjustment; the NPU operating frequency can be set directly in the rknn-smi tool (source: RKNN-SMI Tool User Guide V1.3.1, §6.2).

Viewing and setting the NPU frequency

# 查看当前 NPU 频率和可用频率等级
rknn-smi info -t npu_freq

# 设置 NPU 频率（需从 available 列表中选取)
rknn-smi set -t npu_freq -d 0 -c 0 -s 600

Performance mode switching (EFFICIENT / NORMAL / PERFORMANCE)

The performance mode caps the maximum NPU operating frequency; see the table above for the mappings.

# 切换到 EFFICIENT 模式（NPU 最高 500MHz，SODIMM)
rknn-smi set -t work_mode -s 0

# 切换到 PERFORMANCE 模式（NPU 最高 1000MHz，SODIMM)
rknn-smi set -t work_mode -s 2

Prefill Mode setting

Prefill Mode controls the NPU frequency strategy during the Prefill phase of LLM inference (source: RK182X/RK1820 AI SDK Quick Start V1.0.4, §7.4):

# 查询当前 Prefill 模式
rknn-smi info -t prefill_mode

# 设为 EFFICIENT（降低峰值电流)
rknn-smi set -t prefill_mode -s 0

# 设为 PERFORMANCE（降低首 token 延迟，默认)
rknn-smi set -t prefill_mode -s 1

What is the thermal control strategy?

The RK182X/RK1820 modules use active fan cooling by default, with a software thermal strategy that adjusts fan speed and NPU frequency step by step (source: RK182X/RK1820 AI SDK Quick Start V1.0.4, §7.3).

Default thermal control mapping table

Tmargin (°C)Fan speed ratio (%)Max NPU frequency (MHz)
001000 [1]
46511000 [1]
60631000 [1]
68781000 [1]
761001000 [1]
83100850
90100500
100 [2]——

[1] The RK1820/RK182X M.2 modules are capped at 850 MHz. [2] When the system temperature exceeds 100°C, the chip risks unstable operation and the system immediately triggers a forced reboot to protect the hardware.

Thermal control uses hysteresis logic to avoid frequent level switching:

  • Rising trigger: when the temperature reaches a Tmargin threshold, the corresponding level configuration is applied immediately
  • Falling recovery: when the temperature drops to Tmargin - 7°C, the system automatically switches back to the previous level

Thermal control commands

# 获取当前系统温度
rknn-smi info -t temp

# 获取当前风扇转速
rknn-smi info -t fan

# 获取 NPU 频率信息
rknn-smi info -t npu_freq

# 查看所有热控策略
rknn-smi info -t thermal_all

Custom thermal strategy

# 设置自定义风扇策略（温度 风扇转速比)
rknn-smi set -t thermal_fan -s "40 20, 50 40, 60 60, 70 80, 80 100"

# 恢复默认风扇策略
rknn-smi set -t thermal_fan -s default

# 关闭自动温控（Host 接管控制权)
rknn-smi set -t thermal_fan -s disable

Note: all thermal strategies are stored in DRAM only and revert to default after reboot. Keep thermal_critical at Default to ensure system stability (source: RK182X/RK1820 AI SDK Quick Start V1.0.4, §7.3.2.5).

How do I measure RK182X power?

The RK182X/RK1820 EVB1 and RK3588 EVB10 (SODIMM slot on the screen side) have on-board power sensors that can be queried directly (source: RK182X/RK1820 AI SDK Quick Start V1.0.4, §4.3.1.2):

# 查看所有设备功耗
rknn-smi info -t power

# 实时监控
rknn-smi info -w

Software method (estimation)

# 查看温度信息
rknn-smi info -t temp

# 查看电压信息
rknn-smi info -t voltage

Hardware method (accurate measurement)

  • Insert a shunt resistor (0.01 ohm) in series at the power input
  • Use a current probe + oscilloscope to capture dynamic waveforms
  • Whole-unit power: insert a power meter in series at the power input

Sleep and wakeup

The RK182X supports sleep/wakeup over a PCIe connection (source: RK182X/RK1820 AI SDK Quick Start V1.0.4, §7.2):

# 休眠
rknn-smi suspend -d 0

# 唤醒
rknn-smi wakeup -d 0

Note: only the PCIe connection supports sleep/wakeup; USB and EtherNet connections do not.

Power optimization tips

  1. Pick a suitable performance mode — use EFFICIENT mode for low-load scenarios, lowering the maximum NPU frequency (500MHz) to save power
  2. Prefill Mode optimization — for LLM inference, EFFICIENT Prefill Mode cuts peak current from 28.37A to 15.35A (source: RK182X/RK1820 AI SDK Quick Start V1.0.4, §7.4)
  3. Standby sleep — when unused for long periods, put the device to sleep with rknn-smi suspend
  4. Custom thermal curves — tune fan speed thresholds to your actual cooling conditions, balancing noise and cooling
  5. Power off unused devices — in multi-device setups, idle devices can sleep