The VectorBlox TinBiNN Deep Learning Overlay is an extension for the VectorBlox ORCA RISC V soft-core processor for deep learning applications in very small footprint FPGAs. Available now for the Lattice iCE40 UltraPlus FPGA, deep learning for computer vision is available in less than 5 milliWatts of power!

The TinBiNN Deep Learning Overlay adds both classical vector operations, implemented as lightweight vector extensions or LVE™ for RISC V, and a convolution accelerator for deep neural networks with binary weights. Programming the TinBiNN Deep Learning Overlay is done purely in C/C++, combining high-performance with easy programmability.

The TinBiNN Deep Learning Overlay also includes a convolution accelerator for deep neural networks with binary weights. It computes one or more 3×3 convolutions concurrently every clock cycle, accepting 8-bit input data and producing 16-bit outputs. Convolutions, which are the most common type of filter used for computer vision applications, can be sped up by a factor of 80 using this accelerator.

The LVE™ adds the ability to efficiently stream data through the RISC V ALU when operating on arrays of data. Compatible with the MXP™ processor, the LVE™ instructions operate on data stored in a dedicated scratchpad RAM. It enhances the performance of standard RISC V processors by an order of magnitude in many applications. It is optimized to fit into even the smallest of FPGAs. Users write standard C programs augmented with a set of optimized LVE™ specific extensions. These extensions are easy to learn and represent typical operations like vector-add or vector-multiply. They can be used to accelerate the implementation of activation functions and filters that aren’t directly computable using the convolution accelerator.

The VectorBlox TinBiNN Deep Learning Overlay comes in two configurations, high performance and low power.

  • Accelerates time to market
  • Only need C/C++ programming skills – No RTL or HW skills required
  • Decouples HW and SW design
  • HW speeds delivered in software
  • Up to 80x performance enhancement to standard RISC V implementations
  • Lightweight vector instructions for up to 10x performance
  • Convolution accelerator for up to 80x performance
  • Tightly integrated with VectorBlox ORCA RISC V soft processor
  • Available in to configurations, high-performance and low-power
  • C compiler with LVE™ extensions and libraries
  • Functional simulator; run and debug LVE™ C programs on any computer system
  • Machine Vision
  • Imaging
  • Smart Camera
  • Internet of Things
  • Mobile
  • Unencrypted RTL
  • Functional Simulator
  • C Extensions and Libraries for LVE™
  • User Documentation
  • Sample Code

Available Formats

The VectorBlox TinBiNN Deep Learning Overlay is available in the following configurations:

Description Target FPGA Fabric Scalability
TinBiNN Lattice iCE40 UltraPlus Custom (Wishbone) Fixed size (ask for options)