Here is a bit detail of the neural network chip. [1]
Anything aside, as a former embedded dev, I would never use there kinds of voodoo chinese chips. Don’t get me wrong, it’s not because it’s from China. It is because of __zero__ well-written public documentation.
Just listing specs in datasheets is meaningless. The datasheet is translated but it does nothing to devs because it has absolutely zero about low level peripherals and its core. You cannot imagine how painful writing hardware-level code is when the documentation doesn’t help. I’m 100% sure even the chinese version documentation has almost zero information too.
Look, they have a sort of SDKs and the only example is one-line-long Hello World. [2]
It seems like it quantizes FP16 into INT8. That's a waste, because many models can work perfectly fine with binary/ternary weights, and some can even tolerate low precision activations.
Anything aside, as a former embedded dev, I would never use there kinds of voodoo chinese chips. Don’t get me wrong, it’s not because it’s from China. It is because of __zero__ well-written public documentation.
Just listing specs in datasheets is meaningless. The datasheet is translated but it does nothing to devs because it has absolutely zero about low level peripherals and its core. You cannot imagine how painful writing hardware-level code is when the documentation doesn’t help. I’m 100% sure even the chinese version documentation has almost zero information too.
Look, they have a sort of SDKs and the only example is one-line-long Hello World. [2]
[1]; https://github.com/kendryte/kendryte-doc-ds/blob/master/en/0...
[2]: https://github.com/kendryte/kendryte-freertos-sdk/blob/devel...