OPENEDGES ENLIGHT Classic

It is a 4/8-bit mixed-precision neural processing unit. ENLIGHT Pro is a separate listing.

It states convolution, pooling, and non-linear activation, a compiler that groups layers to reduce DRAM traffic, and 4-/8-bit mixed quantization. Hardware advantages name 4/8-bit computation, a DNN vector engine, and multi-core scale-out. The toolkit names ONNX (PyTorch), TF-Lite, and CFG (Darknet), with quantization from float to 4-/8-bit integer. Applications listed include person, vehicle, bike, and traffic-sign detection, parking-lot vehicle location, license-plate detection, and surveillance tracking. Deliverables are RTL for synthesis, software toolkits and a device driver, a user guide, and an integration guide. Figures stated only in the ENLIGHT Pro section are not repeated here. The Classic section does not name an operating system for its device driver. Use Classic only for the 4/8-bit product described on the NPU page. Specifications that appear only in the Pro section are not recorded for Classic. No process node or foundry is stated.