COMPUTER VISION

Fully Hardwired Deep Learning Inference Object Detection super-resolution HW IP supporting UHD resolution upscaling based on deep learning neural networks.

c.WAVE100

Object Detection HW IP

c.WAVE120

Super Resolution HW IP

 

c.WAVE100

Object Detection HW IP

Chips&Media's Computer Vision IP is Deep Learning based Object Detection with the capability to process 2K resolution at 30 FPS input in real-time. The key differentiating feature of c.WAVE100 Object Detection IP is a fully hardwired and neural network dedicated architecture for a significant reduction in memory accesses and memory bandwidth requirements. Moreover, c.WAVE100 provides high performance with low power consumption, as the IP is optimized to target computing-intensive edge devices in automotive and surveillance applications.​​

View Detailed Specs


Product Specification

  • Application-Specific Neural Networks
Pruning
  • More than half of all weights are zero coefficients
Quantization
  • 8-bit activation, 8-bit bias with a dynamic fixed point per layer
  • Log-quantized weights​​
Network Dedicated Hardware IP 1,168 MACs in FLX (Full Layer Accelerators)
  • Optimized area
  • Multiply-less MAC operation, saving 30% logic gates (compared to typical MAC)
Zero-Skipping
  • Saves processing time and power consumption
  • Skips MAC operation @ IA = zero
  • Saves power @ W = zero
  • Fusing Layers
  • Reduces bandwidth and saves power consumption
  • Saves external memory bandwidth





c.WAVE120

Super Resolution HW IP

c.WAVE120 is a deep learning-based, super-resolution IP that upscales low-resolution data into high-resolution in real-time.

 

c.WAVE120 performs this task by utilizing a massive set of training datasets.  When low-resolution images or videos are zoomed-in, the pixels appear broken and blurry, but that’s when our SR technique steps into action. c.WAVE120 extracts the feature points of an image or video, splits them pixel by pixel, applies the appropriate colors to fill in the missing parts of the data, stitches them, and then reproduces sharper, high-resolution images. c.WAVE120’s neural network was designed and trained to upscale video horizontally and vertically to two,  and four times larger with improved resolution results. For example, it does 4K UHD video to 8K UHD horizontally and vertically, and with a 1080p HD video, it can convert it to the 4K UHD format.

In modern technology, neural networks are used to implement neural processing units (NPU) to execute deep learning algorithms.

For the SR network, implementation is impracticable due to the structural limitations, such as extremely high DRAM bandwidth requirements. This technology can be applied to various customer application product markets such as consumer electronics, automotive, home entertainment, IoT, surveillance cameras, and much more.

View Detailed Specs


Product Specification

  • Performance: 8K (7680x4320) 60fps @550MHz
  • Supported Input Image Format: YUV 400/420 (Optionally YUV 422)
  • The Supported Bit Width of In/Out Image: 8-/10-bit (Optionally 12-bit)
  • Operation Mode: M2M and OTF (On-the-Fly or DRAM less) mode
    • Supporting frame compression and non-frame compression for the input/output pixel data of c.WAVE120
  • Crop Mode: Supporting crop mode for the input image
  • Supported Scaling Ratio of Neural Network: The starting position and the size of crop region should be the multiple of 2
    • x2, x4
  • Supported Scaling Ratio: An arbitrary scaling ratio between x1.0 and x8.0
    • The width of each input and output component (Y, Cb, and Cr) should be the multiple of 4
  • Features Extraction​
  • Constructing HR Image
  • Cost-effective high-quality IP ​





 

 Quick Navigation 

Documentation

Products

Support

  • LinkedIn
  • Facebook
  • YouTube

All contents © Copyright Chips&Media Inc. 2020. All Rights Reserved.