Neowine to develop own brand of AI chip IPs

South Korean chip fabless company Neowine was developing its own AI chip design platform, its CEO said.

Neowine CEO John H. Lee told TheElec that the IP and software platform will be called EONNX, or Embedded ONNX, which will be based on the open source AI software platform Open Neural Network Exchange (ONNX).

Current deep learning frameworks offered by foreign companies have no data compatibility with each other, which means they can only work on certain software and hardware, Lee said.

Neowine’s ONNX platform aims to resolve this issue, the CEO said.

The platform will make Nvidia’s Cuda, Facebook’s PyTorch and Google’s TensorFlow deep learning frameworks compatible with each other, he said.

Developers will be able to apply their AI platform developed in another framework to another one they wish to use. Lee said.

ONNX is a standard and there needs to be a software platform so that it can be embedded on chips, the Neowine CEO said.

The company was also using EONNX to develop its own neural processing unit that has high processing and operating speeds, which was a national project given by the South Korean government, Lee said.

The chip will finish FPGA level test next year and will be realized using the software platform, he added.

Neowine was founded in 2002 and specializes in security chips that prevent copying. Its ALPU chip had sold 120 million units since launch, Lee said. AI chips are being pushed as its new growth engine.