Movidius

From RidgeRun Developer Wiki

Introduction

This wiki contains basic information about Intel Movidius and its VPU mainly designed for deep learning inference. It contains useful information to guide user through the tools provided by Intel and where to find documentation about the chip.

Movidius is looking to enable deep learning at the edge.

Device

Myriad 2

Figure 1.Movidius - Image from Intel Movidius site
  • The chip name is Intel Movidius Myriad 2 VPU and it is available as a USB 3.0 stick.
  • ISP 600 Mpixels per second
  • Movidius proprietary processor is called SHAVE (Streaming Hybrid Architecture Vector Engine)

Neural Compute Stick (NCS)

It also contains the Myriad 2

Figure 1.NCS - Image from Intel Movidius site
  • Myriad 2 VPU is available already
Intel® Movidius™ Myriad™ 2 VPU
12 Programmable SHAVE cores
Low-power design for deep neural networks
6 HD camera support
Configurable ISP pipeline
MIPI 1.1, USB3.0, SPI, I2S, SD, 1GE
1GB RAM, 
4GB free storage space

Myriad X

A new Myriad™ X will be released soon with 16 Shave cores with the following features according to Intel

16 Programmable SHAVE Cores
Neural Compute Engine
8 HD camera support
Configurable ISP pipeline with 4K encode
MIPI 1.2, USB 3.1, Quad SPI, 2x SD, PCI-E Gen3, 10GE, I2S
16 MIPI CSI Lanes
4K @30fps or 60fps
3 pair of stereo 720p

Stages involved

  • Compile: It converts a cafe network into a embedded neural network optimized for Movidius.
  • Tune: Inspect Layer-by-layer performance metrics to improve the network.
  • Accelerate: discrete neural network accelerator for some systems without AI capabilities.

Framework and SDK supported

Figure 1.MDK - Image from Intel Movidius site

Quick Start Guide

Forum

There is a forum (soon moved to Intel® Developer Zone:

https://ncsforum.movidius.com/discussions

Power consumption

  • 100 Giga Flops
  • 1W

Release Date and Price

  • July 2017
  • Price $79

Reference links