RAPTOR - TinyML & Edge AI Platform

Speed up your algorithms at a fraction of the energy spent by your CPU

Technology for better future

The solution to deal with data deluge, while preventing the increase of power consumed by data centers, is known as Edge AI. This solution consists in transferring most of the processing intelligence from the cloud to the sensor. It translates into an unprecedented need to increase performances of «smart devices» by a factor of 1,000 at constant energy consumption.

With our SPEED IP platform, we are positioned as THE provider of solutions for Edge AI System-on-Chip designers. We enable our customers to do much more with less energy resulting in major benefits on environment.

RAPTOR Neural Processing IP Platform

Tiny RAPTOR is a Power efficient Neural Processing IP Platform specialized in sound and vision. Tiny RAPTOR’s near-memory computing architecture is composed of a DMA, local memory and up to 128 processing elements in its most powerful version.

Tiny RAPTOR is a fully programmable accelerator designed to execute deep neural networks (DNN) in an energy-efficient way. It reduces the inference time needed to run Machine Learning (ML) Neural Networks (NN). Tiny RAPTOR fits particularly well within any MCU subsystem, in particular with our MCU subsystem CHAMELEON.


  • Home appliance: door lock, smart speaker, voice controlled system​
  • Image classification: General Purpose and Ultra Low Power MCU
  • Voice controlled equipment: Audio, health, industrial, metering, metrology
  • Connectivity and low power IoT end-nodes: Smartwatches, TWS, Smart Shoes

Key figures

  • 3x higher power efficiency compared to state-of-the-art NPU (KWS-TinyML)
  • ​30mW power @ mobilenet v2 image classification (224x224-60FPS-500MHz)
  • Up to 256GOPS peak at 1GHz
  • 2.2TOPS/W computing efficiency
  • Small footprint (0.045mm2 in 22nm for 32 GOPS)​​

Tiny RAPTOR, a neural processor

Tiny RAPTOR is a specialized neural processor that implements up to 128 processing elements and the necessary control and arithmetic logic to execute Machine Learning Algorithms of predictive models.

Tiny RAPTOR’s scalable architecture is composed of multiple processing branches (up to 8), each of these branches being composed of processing elements (up to 128) aggregated in Neuro Computing Blocks. Each of these block feature contains 4 processing elements, the embedded data memory, and the routing interconnect.

Tiny RAPTOR comes with a complete toolchain running the most popular frameworks of machine learning models AI environments (SDK with Dolphin Design’s specific HAL drivers and documentation package) which complements Chameleon deliverables.

Tiny RAPTOR specific architecture will accelerate the performance of common machine learning tasks and inference sequences for image classification, object detection, keyword spotting and other predictive models.

Key benefits

  • Specialized computing to achieve ultimate energy efficiency
  • Plugin accelerator to enhance the performance of your MCU
  • Hardware flexibility to cover various NN structures
  • Real time image and audio processing
  • Reduced memory footprint with compression
  • Native compatibility with standard AI frameworks

Key features

  • Pre-verified and highly Flexible TCDM for near memory processing
  • Complete SDK (AI Toolchain, ISS, Drivers)
  • Scalable from 32 to 128 MAC/cycle
  • Signed and unsigned 8-bit architecture with SIMD support
  • Lossless weight compression
  • Embedded DMA for blind weight and activation transfer
  • Events manager

Download our white paper “At the edge of data processing”

In order to implement efficient data processing solutions at the edge, MCU architectures need to be modified.
Firstly, an efficient fine-grained data power network needs to be implemented, optimizing not only leakage, but also dynamic power.
Then a new sensor-centric approach must be implemented, to avoid involving the CPU in all events in the case of large data collection.

  1. The need for more and more edge processing capability
  2. Limitations of current MCU solutions
  3. What needs to be changed
  4. Dolphin Design SPEED MCU subsystem and computing platform offer
  5. Benchmarks
  6. Example of audio applications
  7. Take-away

Looking for an AIoT systems to address energy efficiency challenges

Need to discuss about your project?

Would you like to be updated on Dolphin Design’s news, events and offers (I understand I can opt-out at any time – read our privacy policy)