TinyRaptor is a fully-programmable AI accelerator designed to execute deep neural networks (DNN) in an energy-efficient way. TinyRaptor reduces the inference time and power consumption needed to run Machine Learning (ML) Neural Networks (NN) while being scalable and a seamless solution to deploy AI/ML in every SoCs. TinyRaptor is particularly well suited for edge computing applications on embedded platforms with both high-performance and low-power requirements.
Read our news "Dolphin Design wins an Embedded Award for Tiny Raptor, its Energy-Efficient Neural Network AI Accelerator"

NPU evaluation platform >

Key Benefits

  • Near-memory compulting techology to improve energy efficiency and decrease memory bandwidth requirements
  • Hardware flexibility to cover various NN model architectures
  • Native compatibility with standard AI frameworks (Keras, Tensorflow, Pytorch,...)
  • Robust and easy-to-use SDK for seamless programmation of the hardware
  • Easy and swift evaluation of the model performances using TinyRaptor model and virtual platform

Key Performances

  • High energy efficiency >5 TOPs/W
  • More than 90% of energy efficiency as compared to traditional MCU for AI/ML workloads
  • Scalable from 32 to 128 MAC:cycle
  • Extremely small area <0.1 mm²
  • Configurable amount of tightly-coupled memory