VeriSilicon's Cha Kenan: Neural network technology evolves, VeriSilicon will help build open source ML systems

Kenan Cha, Vice President of VeriSilicon Machine Learning Software and Neural Network Processor Products, introduced the VeriSilicon Neural Network Processor at the Khronos·VeriSilicon Technology Seminar.

Cha Kainan said that among all the IP technologies of VeriSilicon, NPU plays a tandem role, and VeriSilicon has great expectations for the intelligent IP market. Neural network technology is constantly developing and evolving, and has a large number of comprehensive applications in computer vision, super-resolution pixel processing, and audio and voice processing. More than 40 companies use VeriSilicon NPU IP technology, and more than 60 products have been mass-produced and commercialized.

VeriSilicon's Neural Processing Unit IP VIP9000 is a highly scalable and programmable computer vision and artificial intelligence processor. VIP9000 uses Vivante's latest VIP V8 NPU architecture. The VIP V8 architecture improves the flexibility of data allocation and the configurability of the processing core to adapt to various filter shapes and sizes in modern neural networks (e.g. 1x1, Nx1, 1xN, depth wise). VIP9000 implements neural network inference through different data formats based on design choices (INT8, INT16, Float16, Bfloat16).

VIP9000 supports all popular deep learning frameworks (TensorFlow, Pytorch, TensorFlow Lite, Caffe, Caffe2, DarkNet, ONNX, NNEF, Keras, etc.) and programming APIs such as OpenCL and OpenVX. In addition, the VIP9000 architecture can also natively support neural network optimization technologies such as quantization, pruning, and model compression.

In the future, VeriSilicon will help build an open source ML ecosystem and provide support for MLIR, IREE, and TVM. Dynamic graph support, SPIR-V integration and support, and support for low-level programming interfaces are also future directions.