RK3588 NPU Shines for Local LLMs: IndieDroid Nova and ODROID-M2 Performance Comparison
RK3588 NPU Excels for Local LLMs: IndieDroid Nova and ODROID-M2 Performance
We’re excited to share that we partnered with security researcher and AI specialist Trevor Unland to provide him with our IndieDroid Nova for his recent tests. Trevor’s excellent
blog post shows how the Rockchip RK3588’s NPU dramatically outperforms the Raspberry Pi 5 when running local LLMs like Llama 3 — with impressive speed and efficiency on full NPU acceleration.
Read Trevor’s original post here: Run Llama 3 Locally: RK3588 NPU vs Raspberry Pi 5
Since several great boards share the same powerful RK3588 processor, we decided to run the same type of tests on another popular option with 16GB RAM: the ODROID-M2. Using a quantized Qwen 2.5 3B Instruct model with the RKLLM toolchain, the ODROID-M2 delivered even stronger results:
- ~8.45 tokens/second sustained on long generations (compared to ~7 t/s on the Nova)
- Lower RAM usage (~2.5 GiB vs 4 GiB)
- Higher NPU utilization (~83% average vs 68%)
- Perfect accuracy and no hallucinations on benchmark tasks
Both the IndieDroid Nova and ODROID-M2 are outstanding platforms for on-device AI in 2026. The Nova offers excellent performance in a compact, feature-rich package, while the ODROID-M2 benefits from Hardkernel’s renowned long-term software support and optimization for the RK3588, along with M.2 NVMe Support for extra and more reliable storage. Whichever you choose, you’ll get fast, efficient local LLM inference without relying on the cloud.
Explore the IndieDroid Nova and ODROID-M2 at ameriDroid today!