Skip to content

Latest commit

 

History

History
47 lines (36 loc) · 2.5 KB

File metadata and controls

47 lines (36 loc) · 2.5 KB
title NPU
description Setting up and using the Neural Processing Unit on Rockchip SoCs with BredOS
published true
date 2026-03-20 09:09:08 UTC
tags npu, rk3588, ai, machine-learning
editor markdown
dateCreated 2026-02-18 09:54:15 UTC

1. Introduction

Some Rockchip SoCs include a dedicated Neural Processing Unit (NPU) designed to accelerate machine learning inference. The RK3588 integrates a 6 TOPS NPU with 3 cores, capable of running quantized neural network models significantly faster than the CPU alone.

There are two separate software stacks for the RK3588 NPU, each with different kernel requirements:

Stack Kernel License Capabilities
Rocket + Teflon (open-source) Mainline 6.18+ GPL / MIT TFLite quantized CNN inference (limited ops)
RKNN-Toolkit2 (proprietary) Vendor BSP only Proprietary Full inference: YOLO, LLM, speech, multimodal
{.dense}

BredOS provides two kernel tracks: Cutting Edge (mainline) and Legacy (Rockchip BSP). The open-source Rocket + Teflon stack works on Cutting Edge kernels 6.18 and later. The proprietary RKNN-Toolkit2 requires a BSP kernel and works on BredOS Legacy images. Note that Cutting Edge (mainline) images are currently available only for a handful of boards — most boards ship with Legacy images by default. {.is-info}

2. Differences between stacks

2.1 When to Use Which Stack

  • The following table gives a brief overview on when to use which stack:
Use Case Recommended Stack BredOS Image
Simple CNN classification (MobileNet) Rocket + Teflon (open-source) Cutting Edge (mainline)
Object detection (YOLO) RKNN-Toolkit2 Legacy (BSP)
LLM inference on NPU RKNN-LLM Legacy (BSP)
Speech recognition RKNN-Toolkit2 Legacy (BSP)
Long-term mainline support Rocket + Teflon (improving upstream) Cutting Edge (mainline)
{.dense}

If you need maximum NPU performance or support for complex models (YOLO, LLMs, transformers), the RKNN-Toolkit2 with a vendor BSP kernel is currently the more capable option. The open-source Rocket + Teflon stack is actively improving and is the recommended long-term path for mainline kernel users. {.is-info}

2.2 Further Information

To find further information on how to use the NPU of your chip refer to the article according to your driver.