Hardware Installation & Verification

Overall block diagram

RK3588 / RK3576 host board        RK182X (PCIe AI co-processor)
───────────────────────           ───────────────────────
PCIe/M.2 slot  ◄────────────►     No application CPU
Power supply (independent)  ◄────► Independent power cable
       │
       ▼
  Debian 12 aarch64
  /dev/pcie-rkep-<BDF>
  rknn3.service / rknn3_transfer_proxy

1. Prerequisites

  • RK3588 / RK3576 host development board + RK182X AI co-processor
  • System booted (Debian 12 bookworm / aarch64)

2. Log In to the Host

Pick any one of the three methods.

Method 1: HDMI + Keyboard & Mouse (Simplest)

Connect a display + USB keyboard and mouse; open a terminal after boot.

Method 2: ADB Connection (USB)

Connect the computer to the board's OTG port with a USB data cable:

adb shell

Default user is linaro.

ADB connection

1. Find the board's IP

Run on the development board (with a display attached, or check the router's admin page):

ip a | grep inet

Find an IP on the same subnet, e.g. 192.168.1.100.

Without a display, look up the device with hostname linaro in the router's admin page.

2. SSH login

Run on the computer:

ssh linaro@<board-ip>

Default username linaro, password linaro.

SSH login

On the first connection you will be asked whether to trust the host; enter yes.

3. Verify

whoami        # should output linaro
uname -a      # confirm the system is aarch64 / Debian

3. Verify Hardware Detection

Check the PCIe Device

lspci | grep Rockchip

Expected output:

0004:41:00.0 Processing accelerators: Rockchip Electronics Co., Ltd. Device 182a

Notes:

  • The device ID usually shows as 182a (the PCIe-side device ID of the 1828 chip); this is normal
  • The BDF address (e.g. 0004:41:00.0) varies with the board's PCIe topology; seeing Rockchip's Processing accelerators entry is normal

One-Shot Health Check

cat <<'EOF' > /tmp/check_1828.sh
#!/bin/bash
echo "=== RK1828 健康检查 ==="
# 自动探测 BDF（实际 BDF 因主板 PCIe 拓扑而异，如 0004:41:00.0）
bdf=$(ls /dev/pcie-rkep-* 2>/dev/null | head -1 | sed 's|/dev/pcie-rkep-||')
[ -n "$bdf" ] && echo "✅ 设备节点 OK (${bdf})" || echo "❌ 设备节点缺失"
systemctl is-active rknn3.service >/dev/null && echo "✅ 服务运行中" || echo "❌ 服务未运行"
# 进程名可能带 hash 后缀（如 rknn3_transfer_proxy_c36211b1），也可能无后缀，故放宽匹配
npu_pid=$(pgrep -f rknn3_transfer_proxy | head -1)
if [ -n "$npu_pid" ]; then
    echo "✅ 代理进程 PID=$npu_pid, CPU=$(ps -p $npu_pid -o pcpu= | tr -d ' ')%"
else
    echo "❌ 代理进程不在"
fi
/usr/bin/pcie_upgrade_tool ld 2>&1 | grep -q 182a && echo "✅ PCIe 链路正常" || echo "❌ PCIe 链路异常"
EOF
chmod +x /tmp/check_1828.sh && /tmp/check_1828.sh

Expected output:

=== RK1828 健康检查 ===
✅ 设备节点 OK (0004:41:00.0)
✅ 服务运行中
✅ 代理进程 PID=1952, CPU=25%
✅ PCIe 链路正常

The script checks 4 things one by one:

  • Whether the /dev/pcie-rkep-* device node exists
  • Whether the rknn3.service service is active
  • Whether the rknn3_transfer_proxy daemon is running
  • Whether pcie_upgrade_tool ld can detect 182a on the link

4. Troubleshooting

SymptomCauseFix
No Rockchip in lspciCard not seated / insufficient power / wrong slot modeReseat + check power + dmesg | grep -iE 'pcie|pcie-rkep|182a|fw_load'
Device node missing (/dev/pcie-rkep-*)PCIe not enumeratedReboot the host or sudo modprobe pcie_rkep
rknn3.service not activesystemd service errorsudo systemctl restart rknn3.service
Daemon rknn3_transfer_proxy missingRuntime did not startsudo systemctl restart rknn3.service
Daemon CPU usage abnormally highInference task stuckrmmod pcie_rkep && modprobe pcie_rkep && systemctl restart rknn3.service
Inference task failsDriver errorSame as above + check PCIe errors in dmesg

5. Next Steps