NPU Integrated Computing


Introduction


  NPU integrated computing solution — a CPU (RK3399) and multiple NPUs (RK1808) combine to an AI cluster with ultra-high computing power and simultaneous AI computing capability. The solution can not only increase the computing power, but also control the costs flexibly with higher cost performance. 


image.png


Example



  With one RK3399 as the master board and eight RK1808, it can simultaneously perform object recognition and analysis in 8-channel videos on Ubuntu system.


Demo

Features


(1) Equipped with AI SoC — RK1808

The built-in NPU computing power of RK1808 can reach up to 3.0 TOPs. It supports INT8/INT16/FP16 mixed precision computing, and balances performance, power consumption, precision to the greatest extent. Besides, it supports network model conversion of TensorFlow, Caffe, ONNX, Darknet and other frameworks. SDK , tutorials and other technical documents are provided.

image.png

(2) Flexible combination with high cost performance

  One RK3399 can be combined with multiple RK1808 to form a high-performance AI acceleration cluster — NPU computing power can be increased to 1+N, and the computing efficiency will be greatly improved. The number of RK1808 can be flexibly adjusted according to the usage scenarios — costs are effectively controlled. 

image.png