VeriSilicon announces VIP9000, a highly scalable and programmable processor for computer vision and artificial intelligence. The Vivante VIP family’s patented Neural Network (NN) engine and Tensor Processing Fabric technology delivers superb neural network inference performance with industry-leading power efficiency (TOPS/W) and area efficiency (mm2/W), with scalable compute capability ranging from 0.5TOPS (Tera-Operations-Per-Second) to 100s of TOPS.
VIP9000 adopts Vivante’s latest VIP V8 NPU architecture. According to VeriSilicon’s Executive Vice President and GM of Intellectual Property Division Wei-Jin Dai, VIP V8 architecture improves the flexibility of data distribution and processing core configurability to adapt to a wide range of filter shapes and sizes in modern neural networks (e.g. 1×1, Nx1, 1xN, depth wise). VIP9000 enables neural network inference with different data formats based on design choice (INT8, INT16, Float16, Bfloat16). VIP9000 also supports hybrid quantization (mixing data formats between neural network operations) natively.
Industries with AI Vision, AI Voice, AI Pixel, or AIOT applications will benefit from VIP9000. For smart home and AIOT applications, VIP9000 offers several highly optimized, high precision recognition engines. The release contains the following new features:
- A more flexible data distributor and processing core configurator: Brings high MAC utilization to a wide range of filter shapes and sizes in modern neural network models;
- New data format support for Bfloat16: On top of existing INT8, INT16, and Float16 support, Bfloat16 delivers better accuracy for AI training;
- FLEXA API support: A hardware and software protocol that enables efficient data communication between multiple pixel processing IP blocks. Systems using VeriSilicon’s IPS, Video CODEC, NPU, or 3rd party IP compliant with FLEXA API can run AI applications with reduced DDR traffics and low pixel processing latency for applications running thru multiple IPS;
- Task-specific engines designed for speeding up commonly used AI applications: This allows for face detection, face recognition, facial landmark detection, object detection and AI Voice. One or more engines can run in parallel inside VIP9000 together with user-defined AI programs, due to VIP9000’s native multi-task, multi-context support.
VIP9000 supports all popular deep learning frameworks (TensorFlow, Pytorch, TensorFlow Lite, Caffe, Caffe2, DarkNet, ONNX, NNEF, Keras, etc.) as well as programming APIs like OpenCL and OpenVX. Neural network optimization techniques such as quantization, pruning, and model compression are also supported natively with VIP9000 architecture. AI applications can be easily port to VIP9000 platforms through offline conversion by Vivante ACUITYTM SDK, or through run-time interpretation with Android NN, NN API, or ARM NN.