K210 AI Accelerator a Compact Raspberry Pi HAT for Computer Vision Applications

March 10, 2021

Story

XaLogics’s AI Accelerator with K210 SoC comes with a dual-core RISC-V AI processor featuring low power consumption than its competing Coral USB Accelerator, and Intel Neural Compute Stick 2.

It is a cost-effective solution for beginners, as it is equipped with pre-trained ML models with open-source software. Hence, it is expected that it will have great community support. The AI accelerator has a compact form factor similar to that of Raspberry Pi Zero, hence making it flexible to work it various other boards with a 40-pin connector.

K210 AI Accelerator is an add-on for Raspberry Pi. The compact Raspberry Pi HAT uses the Kendryte K210 AI processor that can perform 0.5 TOPs (Tera Operations Per Second) for deep learning computations. The SoC features a 64-bit dual-core RISC-V processor with FPUs. The SHA256 and 128-bit AES accelerator provide on-chip security.

K210 AI Accelerator comes with pre-trained models that can be used to add machine vision features to RPi-based cameras very quickly. It also supports a plugin module that can add AI features to the camera using Python APIs. This reduces the tedious process of training neural networks. The pre-trained models include customizable object detection, face detection, age, and gender estimation, simple voice commands, and vibration abnormally detection.

 

 

 

If you don’t want to use pre-trained models, then you can train your own neural networks separately, for example by using Nvidia GPU. The company uses TensorFlow for training and will provide an example of how to train your own object detection model.  K210 AI Accelerator supports Kendryte KModel conversion tool compatible with TFLite, Caffe, and ONNX format.

Secured Software Support of K210 AI Accelerator

Using the Visual Studio Code and the required toolchain for K210, the development of K210 firmware can be done on Raspberry Pi. The K210 AI Accelerator features Infineon Trust-M as its cloud security chip. This enables a secure connection to AWS via MQTT without revealing the private key. The protocols used to connect the IoT devices to the cloud, provide every IoT device its own unique identity. This pre-personalized solution allows secured and centralized configuration needed for quick cloud access.

K210’s schematics C code and the programs running on “Raspberry Pi will be open-sourced.” Pre-trained models discussed earlier exist in binary form. The form factor of the device is similar to Raspberry Pi Zero HAT. The company had launched the XAPIZ3500 in a similar form factor in the previous year. 

 

Conclusion

You can find a comparison table between K210 AI Accelerator, Coral USB Accelerator, and Intel Neural Compute Stick 2 on the Crowd Supply page. It is clear that the K210 Ai Accelerator is less expensive and features a low power consumption rating of 0.3 W.

However, the NPU performance is a bit on the lesser side in comparison with others but the overall features compensate for this factor.

The K210 AI accelerator can be preordered at $38 on Crowd Supply. However, the shipments will take place at the end of May 2021. The US shipping charges are $8, whereas the for rest of the world it is $16. For more information visit the product page.