OPENEDGES ENLIGHT Pro

It is a neural processing unit for inference. ENLIGHT Classic is a separate listing, and Pro's specifications are not Classic's.

It states Transformer model support, 4,096 INT8 MACs per cycle, and performance scalable from 8 TOPS to hundreds of TOPS. Single-, dual-, and quad-core configurations are listed. Hardware advantages name INT8, INT16, and FP16 mixed precision, a DNN vector engine, and a RISC-V CPU vector extension with custom instructions. The SDK names ONNX (PyTorch), TFLite, and CFG (Darknet). Applications listed include object detection and tracking, face detection, pose and gesture recognition, vision defect inspection, and a natural language interface. Deliverables include Verilog RTL, an example testbench, synthesis constraints, a Linux device driver and API manual, and integration and toolkit guides. The page does not map the TOPS range to a core count. Confirm precision, core count, and the TOPS point for the configuration. The Pro toolkit section also shows a 4-/8-bit quantizer line; that line is not recorded as Pro precision. No process node or foundry is stated.