Paddle-Lite is part of Baidu’s deep learning framework PaddlePaddle (Parallel Distributed Deep Learning), and the IMGDNN API will enable developers to target PowerVR-architecture-based graphics processing units (GPUs) and neural network accelerators (NNAs) in order to improve the performance of their applications.
The NNA has been integrated into the Paddle-Lite framework and increases the number of frameworks supported by Imagination’s hardware platforms and will enable developers to run AI applications inside heterogeneous systems, such as those containing CPUs, GPUs and NNAs.
For example, the ‘Imagination NNA + CPU’ heterogeneous computing system can be established so developers can make the best use of the NNA and CPU.
Developers will be able to take advantage of the complete PaddlePaddle toolset, including PaddleSlim, a compression tool for optimising pre-trained neural network models and converting them so they are suitable for execution on Imagination’s NNA. The code has already been merged into the PaddlePaddle’s development branch and will appear in the next major stable release V2.8.
Yunkai Wang, Ecosystem Product Owner, PaddlePaddle, said, "It is exciting to be collaborating with Imagination in the AI space. PaddlePaddle’s Paddle-Lite inferencing engine can support multiple types of hardware, operating systems, and AI models comprehensively and the successful integration with Imagination’s hardware enriches our ecosystem further."
Imagination’s NNA IP is designed to accelerate machine learning workloads at the edge. It is silicon-proven and has been licensed into markets including automotive, mobile, AIoT and datacentre/desktop. Featuring ultra-high performance per mm2 the new IMG Series4 multi-core is power-efficient and can scale to 500 TOPS and beyond. Architectural highlights include Imagination Tensor Tiling to reduce bandwidth needs as well as mature software and tools.
Commenting Andrew Grant, Senior Director, Artificial Intelligence, Imagination Technologies, said "Imagination NNA can support all mainstream AI frameworks at present and the support for Paddle Lite engine means our ability in this area continues to strengthen."