Movidius Myriad 2

From Virtual Reality, Augmented Reality Wiki
Jump to: navigation, search

Intro

Machine vision industry leader, Movidius, puts low power and high performance together in its latest vision processing unit named Myriad 2. It is aimed at providing intelligence to vision processing on a dedicated SoC, and deliver an extra edge in performance to existing CPU and GPU used in VR hardware.

Priced at $10 per unit, and targeting OEM VR hardware manufacturers, it will deliver features that have never been seen before in devices like smart glasses and Virtual HMD. Motorola has recently announced their selection of Movidius Myriad 2 to power their latest Motor Mod: 360° Camera.

Features

The Movidius Myriad 2 is able to perform a staggering 2 trillion of 16-bit operations per second, while only consuming 500 mW. That, and its ability to process visual data more intelligently is attributed to its brilliant SoC architecture design.

Key in reducing its power consumption is the hardware accelerators called SIPP filters, which also plays an important part in to run pre-configured vision processing tasks such as fusing data from different cameras.

The SoC hosts an array of 12 SHAVE Vector processors, the heart of the unparalleled vision processing power of Myriad 2. Each of the processors runs at 600Mhz and handles code branches efficiently.

The Myriad 2 also features two RISC processor, with one responsible for scheduling within the SoC and another runs the user codes within the Real Time Operating System(RTOS).

Hardware

The Movidius Myriad 2 is offered in two different packages, 225 Ball BGA and 270 Ball BGA with the latter supporting up to 4Gb of DDR 3 RAM. It is built with 2Mb On-Chip memory and 256 KB of L2 cache.

The Myriad 2 SoC supports a rich set of interfaces, that includes I2C, SPI, I2S, USB 3.0, GPIO,1 Gbit Ethernet and a debug interface.

Technical Specifications

  • Heterogeneous, high throughput, multi-core architecture based on
  1. 12 VLIW 128-bit vector SHAVE Processors optimized for machine vision
  2. Configurable hardware accelerators for image and vision processing, with line-buffers enabling zero local memory access ISP mode
  3. 2 x 32-bit RISC processors
  4. Supports data and task parallelism
  5. Programmable Interconnect
  • Support for 16/32-bit floating point and 8/16/32-bit integer operations
  • Homogeneous, centralized memory architecture; 2MB of on-chip memory
  • 400 GB/sec of sustained internal memory bandwidth
  • 256 KB of L2 Cache
  • Power management: 20 power islands; low power states
  • Nominal 600 MHz operation at 0.9 V
  • Rich set of interfaces:
  1. 12 Lanes MIPI, 1.5 Gbps per lane configurable as CSI-2 or DSI
  2. I2C, SPI for control and configuration
  3. I2S for audio input
  4. Banks of configurable GPIO, PWM
  5. USB3 with integrated PHY
  6. 2-Slot SDIO
  7. Debug interface
  8. 1 Gbit Ethernet
  • Available package configurations
  1. MA2150/MA2155: 6.5mm x 6.5mm, 0.4mm pitch, 225 Ball BGA, 1Gb LPDDR II
  2. MA2450/MA2455: 8mm x 9.5mm ,0.5mm pitch, 270 Ball BGA, 4Gb LPDDR III
  • Advanced low-power 28nm HPC process node

Developer

Extensive supports and resources such as the SDK, libraries, framework and camera API are provided to developers.

https://www.movidius.com/solutions/software-development-kit

Background

Founded by Sean Mitchell and Dr. David Moloney in 2005, Movidius held to the belief that hat machine vision applications can only be effective at the network edge, running right beside the sensors providing the input. The Myriad 2, a low power high-performance vision processing unit was born from the lack of a suitable single chip solution that satisfies the vision of Movidius.

References