Neural processor


Architected from ground up for neural processing, RealSilicon's Neural Processor architecture is extremely low power and highly efficient. It is also highly scalable, configurable and easily programmable. When implemented with RealSilicon's physical IP, this provides substantial PPA (performance, power, area) advantage over any other alternative available in industry.

Highlights


  • AD-SIMD Architecture
    • Asynchronous Distributed – Single Instruction
      Multiple Data
  • Efficient SRAM based Array processor
    • Scalable and configurable at design (Area scaling)
    • Also configurable in application (power and performance scaling)
  • Can implement DNN/CNN as well as other
    DSP, filtering tasks
  • AD-SIMD Architecture
    • Software stack is ready
    • Programming using C++
    • Compile using off-the-shelf compiler (gcc)
  • Versatile usage potential
    • Usable in low end applications in IoT such as
      • Voice activation
      • Smart sensor mechanism
    • Usable in high end applications such as


Features


  • Flexibility

  • Doesn’t require new compilerdevelopment
    Re-usable as an IP block for anysuitable host
    Datapaths can be customized forvarious applications

  • Any width, custom datatypes etc.

  • Compatibility

  • Programming model is standard C++ (with custom-instruction
    macros in object methods)
  • No API dependencies

  • Any operating system and host environment
  • Host runs existing algorithms (e.g.OpenCV)
  • Applications are portable to any other platform (e.g. Windows)
  • Orthogonal support of other parallel techniques (e.g. SMP, multithreading)


Differentiations

  • Very high perfomance
    • 100's of processing elements
  • Very low power – Very low power built into architecture Highly scalable
    • Scalable at design
      • Modular design consisting of Neural tiles.
      • Number of tiles can vary from 1 to 1000
    • Scalable in application
      • Power consumption scales with performance
    • Linear scaling - Performance, resolution, frame rate, cost, power, battery life
  • Ease of programming
    • Inherently, efficiently exploit parallelism
    • Completely abstract resource allocation, dataflow, coherency, scheduling, interlocks
  • Software-compatible roadmap
    • Including binary‐compatible

Applications

  • DNN/CNN
  • Object detection
  • Gesture recognition
  • Location-based services
  • Untethered robotics
  • Security
  • Automated image/video search
  • Sensor enhancement, display augmentation

Contact Us



Main US Office
19925 Stevens Creek Boulevard
Suite 100
Cupertino, CA USA. 95014
Phone: +1 408 275 2661
info@realsilicon.com


Bangalore Office
Ramky House, Site No. 25-30,
2nd Cross Ragavendra Nagar, Hennur Ring Road, Kalyan Nagar Post
Bengaluru, Karnataka 560043, India
Phone: +91 80 4097 3677
info@realsilicon.com





Contact us to learn more


Get more info