The Kinara AI Software Development Kit (SDK) optimally targets trained AI models onto Ara-1 and Ara-2 silicon and modules. Kinara’s fully programmable compute engines allow our model compiler to quickly adapt to a virtually unlimited range of neural network architectures

Benefits of running foundation models at the source of data collection

  • Creates optimal execution plan

    AI compiler automatically determines the most efficient data and compute flow for any AI graph

  • Extensible compiler

    Converts and schedules models ranging from CNNs to complex vision transformers and Generative AI

  • Readily support new operators

    Fully programmable compute engines with a neural-optimized instruction set

  • Support for multiple datatypes

    INT8 (Ara-1, Ara-2), INT4 and MSFP16 (Ara-2 only)

  • Efficient dataflow for any network architecture type

    Software defined Tensor partitioning and routing optimized for dataflow

  • Utilizes flexible quantization methods

    Choose between Kinara integrated quantizer or TensorFlow Lite and PyTorch pre-quantized networks (Ara-2 only)