Hello World

The RK1828 has no application CPU of its own — programs run on the RK3588 (aarch64). This article covers the first program: write the code, compile it, run it. NPU invocation comes in the last section; first get the pipeline working.

1. Write the code

Create a file hello_rk1828.c:

// hello_rk1828.c
#include <stdio.h>

int main() {
    printf("Hello, RK1828!\n");
    printf("Running on Rockchip SoC.\n");
    return 0;
}

2. Compile

Compile directly on the RK3588.

# 1. Check whether the compiler is installed
which aarch64-linux-gnu-gcc && aarch64-linux-gnu-gcc --version | head -2

Expected:

/usr/bin/aarch64-linux-gnu-gcc
aarch64-linux-gnu-gcc (Debian 12.2.0-14+deb12u1) 12.2.0
# 2. Compile
aarch64-linux-gnu-gcc hello_rk1828.c -o hello_rk1828

# 3. Verify the build result
file hello_rk1828

Output:

hello_rk1828: ELF 64-bit LSB pie executable, ARM aarch64, version 1 (SYSV), dynamically linked, interpreter /lib/ld-linux-aarch64.so.1, BuildID[sha1]=537c84f9fc03eabc1612bee7d06d8c9fc209e87d, for GNU/Linux 3.7.0, not stripped

3. Run it

./hello_rk1828

Output:

hello World program output

4. Deploy to another device (optional)

Only needed when pushing the program to a different board.

scp / ssh

scp hello_rk1828 root@<device-ip>:/tmp/
ssh root@<device-ip> "/tmp/hello_rk1828"

USB drive / SD card

cp hello_rk1828 /media/<user>/USBDRIVE/
# On the target board, run it after mounting; use the actual mount point path

5. A quick look at the board environment

# Kernel
uname -a

Output:

Linux linaro-alip 6.1.141 #27 SMP Wed May 27 13:44:15 CST 2026 aarch64 GNU/Linux
# CPU
cat /proc/cpuinfo | head -5

Output:

processor	: 0
BogoMIPS	: 48.00
Features	: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm lrcpc dcpop asimddp
CPU implementer	: 0x41
CPU architecture: 8
# Memory
free -h

Output (measured):

               total        used        free      shared  buff/cache   available
内存：         3.8Gi       3.2Gi       211Mi       3.8Mi       444Mi       603Mi
交换：            0B          0B          0B

6. Call the RK1828 NPU

The RK1828 has no application CPU of its own; to run NPU inference you must call the rkllm3-server service over PCIe.

# Find where the service is installed
which rkllm3-server

Output:

/usr/bin/rkllm3-server
# Show help
/usr/bin/rkllm3-server --help 2>&1 | head -10

Output (excerpt):

version: 1.0.0 (b4bd144@2026-01-22T21:27:27)

----- common params -----

-c,    --ctx-size N                     size of the prompt context (default: 4096, 0 = loaded from model)
-n,    --predict, --n-predict N         number of tokens to predict (default: -1, -1 = context size)
-a,    --alias STRING                   set alias for model name (to be used by REST API)
-m,    --model FNAME                    rknn llm model path

Start it:

# Start the NPU inference service
rkllm3-server -m <model.rknn> -c 0xff -n 4096

Parameter quick reference:

ParameterMeaning
-m / --modelModel file path (required)
-c / --ctx-sizeNPU core mask; use 0xff for 8 cores
-n / --predictMax tokens to generate, default -1

7. FAQ

SymptomCauseFix
cannot execute binary fileArchitecture mismatchMake sure it was built with aarch64-linux-gnu-gcc
Runs but the NPU does nothingWrong device node addressedThe RK1828 device node is /dev/dri/renderD128
rkllm3-server crashes after a whileResidual KV cacheRestart the service

8. Known issues

  • Crash after consecutive inference: after several runs, rkllm3-server may crash due to residual KV cache; restart the service
  • Network access: when the board cannot reach GitHub / HuggingFace directly, use mirror sites such as ghfast.top, hf-mirror.com
  • Development environment: Node.js ≥ v22 required (v25 via nvm recommended)