SCAMP-5: Vision Sensor with Pixel Parallel SIMD Processor Array

Event: CVPR 2019 Workshop · Duration: 24 min · ▶ Watch on YouTube

Abstract

The talk introduces SCAMP-5, a vision sensor with a 256x256 pixel-parallel SIMD processor array, designed to overcome the data movement bottleneck and power limitations of conventional embedded machine vision systems. By integrating processing directly into each pixel, SCAMP-5 enables high-speed, low-latency, and low-power computation of various early vision tasks and complex information extraction. The speaker demonstrates the chip’s capabilities for algorithms like edge detection, median filtering, optical flow, corner detection, object tracking at 100,000 fps, microfluidics particle tracking, HDR sensing, depth from focus, and even convolutional neural networks, all processed directly on the sensor.

Speakers

  • Piotr Dudek — The University of Manchester

Talks (1)

  • 00:00 — Piotr Dudek: SCAMP-5: Vision Sensor with Pixel Parallel SIMD Processor Array
    • Piotr Dudek introduces the SCAMP-5 vision chip, a pixel-parallel SIMD processor array designed for high-speed, low-power, and low-latency embedded machine vision by performing computations directly on the sensor.

Key Takeaways

  • Integrating processing directly into the vision sensor (vision chip) significantly reduces data movement, leading to higher performance, lower latency, and drastically reduced power consumption for embedded machine vision.
  • The SCAMP-5 chip features a 256x256 array of 65,536 software-programmable processors, allowing for flexible implementation of diverse early vision and more complex algorithms directly on the sensor.
  • The mixed-signal analog/digital datapath and pixel-parallel architecture enable operations like edge detection, median filtering, optical flow, and object tracking at extremely high frame rates (up to 100,000 fps) with very low power consumption (nW/pixel).
  • The system supports ‘sub-frame’ computation for advanced sensing techniques like High Dynamic Range (HDR) and Depth from Focus, which are challenging for conventional sensors due to data bandwidth limitations.
  • Despite being built on 20-year-old CMOS technology (180nm), SCAMP-5 achieves impressive efficiency (535 GOPS/W), highlighting the potential for even greater performance with modern fabrication processes.

Methods / Models / Datasets Mentioned

  • SCAMP-5
  • DVS
  • DAVIS
  • ATIS
  • Celex
  • FAST algorithm
  • L1-norm
  • Ternary Convolution Layers
  • Max pooling Layers
  • SCAMP
  • SCAMP-2
  • SCAMP-3
  • ASPA
  • ASPA/2
  • ASPA3
  • PAV3D
  • MPAC
  • MIMD
  • T3D
  • SCAMP-6
  • SCAMP-7

Topics

Embedded machine vision · Vision chip · Pixel-parallel processing · SIMD processor array · In-sensor computation · Low-power vision · High-speed vision · Event cameras · Agile robotics · Convolutional neural networks

Notes

Open for commentary — connections to other work, critiques, follow-up reading.