NeuVerge Inference Framework

 

White Paper

The NeuVerge Inference Framework is our very own ML model built upon state-of-the-art convolutional neural networks (CNNs). It can be utilized to drive various downstream tasks by compressing raw pixel data into a Feature Map (2D matrix of neurons). The NeuVerge Framework is capable of performing complex tasks in various domains spanning from Computational Vision, Object Detection & Segmentation, and Image Encoding

Parallel Implementation 

  • Convolution neural network (CNN) models implemented in Verilog for parallel evaluation
  • Allows processing modern image sensor data in hundreds of millions of pixel/s with ML model inference
  • Achieving 3 degrees of parallel implementation and producing 10K+ MAC units running in parallel

Application – Real Time Image Encoder IP’s performance 

  • Real-time performance at maximum image sensor resolution and frame rate
  • Supports fusion of multiple input frames into a single output feature map
  • Achieved >100X compression converting from raw pixels to features & can drive multiple downstream tasks
  • Large models can be ASIC hardened for higher clock rate & ROM based weights

ML Model Training Strategies

FPGA – Continuous learning with reconfigurable FPGA bitstreams, enabling the model to learn an unknown distribution over time.

ASIC – Collect a large dataset and train a foundation model with static weights and hardened ASIC implementation. Leading to the highest performance and smallest area, suited for high volume applications