ENLIGHT is a deep learning accelerator designed to enhance inferencing computation with significant efficiency and compute density. It is engineered for edge computing environments, supporting mixed-precision (4/8-bit) operations that reduce DRAM traffic through clever layer partitioning and scheduling. Ideal for various deep neural networks, ENLIGHT delivers impressive performance and energy efficiency, outperforming many alternative solutions.
The IP is highly versatile, suited for a range of applications from vehicle detection to surveillance. It includes a sophisticated vector engine optimized for modern deep neural network algorithms, like depth-wise convolution and feature pyramid networks. The ENLIGHT platform can scale out with multi-core support to parallelize DNN layers, ensuring high performance through efficient inter-layer optimization.
With a comprehensive toolkit that supports industry-standard network formats like ONNX and TF-Lite, ENLIGHT simplifies the integration process for clients, helping them tailor solutions that precisely fit their needs. The design's extensibility allows for easy adaptation to new DNN changes, ensuring clients stay ahead in the rapidly evolving AI landscape.
