Orange Pi AI Stick 2801 Neural Compute Stick SDK & Documentation Released

We covered the launch of Orange Pi AI Stick 2801 neural compute stick a few days ago, the first easily available Gyralcon Lightspeeur based USB stick, and noted that while the hardware was there, we had no details about software development kit and documentation at the time.

I’ve got some more information now. First, the company release an English presentation about the neural stick, and while it’s not technical documentation, it provides insights into what it is capable of, and an overview about the workflow.

How to use Orange Pi AI Stick SDK
Click to Enlarge

More importantly, Shenzhen Xunlong Software also released Orange Pi AI Stick SDK and user manual for x86 PC, as well as cat/dog data and pre-trained models to get started.

However, if you want to perform training with your own data, you’ll need to purchase the stick plus PLAI model transformation and training tools sold on Aliexpress for $218 plus shipping, that’s $69 for the hardware, and $149 for the software tools. The company further explains that the tools support model decomposition of convolutional neural network (CNN) based on Caffe, and the training tool can train image, video, voice, and natural language into prototype to be loaded into the neural stick.

PLAI stands for People Learning Artificial Intelligence , and is an ecosystem of hardware and tools provided directly by Gyrfalcon, the ASIC vendor. Shenzhen Xunlong did not provide the full details about the tools, but the tools should be Plai Builder, a PyTorch based “full-stack” framework specially designed to quickly train and build neural network models for Lightspeeur and G.A.I.N series solutions from GTI (Gyrfalcon Technology Inc).

Orange Pi AI Stick Laptop RK3399 BoardLast time, various people asked what such stick could do, and the generic answer is inference (e.g. object recognition) at the edge at low power, but the presentation also includes a list of potential applications:

  • Moving edge calculation
  • Intelligent monitoring
  • Smart toys and robots
  • Smart home
  • Virtual reality and augmented reality
  • Face detection and recognition
  • Speech Recognition
  • Natural language processing
  • Embedded deep learning device
  • Cloud Machine Learning and Deep Learning System
  • Artificial intelligence data center server
  • Advanced assisted driving and autonomous driving

So lots of fun applications. If you are specially interested in Tensorflow, you’ll have to wait a little longer, as I understand support for the framework is still being worked on.

Share this:

Support CNX Software! Donate via PayPal or cryptocurrencies, become a Patron on Patreon, or buy review samples

Subscribe
Notify of
guest
The comment form collects your name, email and content to allow us keep track of the comments placed on the website. Please read and accept our website Terms and Privacy Policy to post a comment.
10 Comments
oldest
newest
Joe
Joe
2 years ago

Thanks for the update! It looks a little disappointing because you would really need the software, and together it’s too expensive. I’ll wait for the Movidius 2 release and more info on it, it should be much more competitive because it comes with proper software. Otherwise the RK3399Pro is still in the pipeline.

Vincent Bellet
Vincent Bellet
2 years ago

Movidius NCS 2 is “only” 1 TOPS. This stick is 2.8 TOPS …

Joe
Joe
2 years ago

But it’s not clear if we’re talking about the same thing. We’d need objective benchmarks for that. I’m not sure, maybe Intel is talking about int8 inference and this one about int4 inference. Then you’d have to halve the numbers.

Joe
Joe
2 years ago

I have seen several sources now speak of 4 TOPS for the Movidius NCS2, like here the official distributor
https://www.mouser.com/m_new/Intel/intel-neural-compute-stick-2/
The 1 TOPS according to this article refers to the Neural network engine itself, but there are also the Shave cores and additional images processing capabilities.
https://www.electronicdesign.com/embedded-revolution/plugging-intel-s-neural-compute-stick-2
I think there need to be some real benchmarks to say in which application which one is better.

Lalith
Lalith
2 years ago

Thank you for the info about this very interesting device.

Jon Smirl
2 years ago

I’m hoping Allwinner sees this and releases similar tools for the V5 AI engine.

Coldfish
Coldfish
2 years ago

Lindenis folks spot your post but not Allwinner folks.

Vincent Bellet
Vincent Bellet
2 years ago

Let’s hope so!

paul
paul
1 year ago

hi
this model support tensorflow lite? i need an example code to test my 2801 please advise.

thanks
PP

Advertisement